Government transparency frameworks such as the Freedom of Information Act (FOIA) in the United States must balance the public's right to know with a number of other considerations. This paper focuses on one such issue, assessment of whether the deliberative process privilege applies in specific cases under FOIA Exemption 5. Providing automated support to the reviewers charged with making such determinations could help to improve responsiveness while controlling review costs. This paper applies ChatGPT-3.5 to explore three ways in which the emerging family of Large Language Models (LLM) might help reviewers with this task: (1) suggesting which passages should and should not be withheld, (2) explaining the basis for those suggestions to the reviewer, and (3) helping the reviewer explain the basis for their decisions to the requestor. The results show that suggestions by ChatGPT-3.5 are not more accurate than previously reported supervised text classification results, that legal analyses in explanations provided by ChatGPT-3.5 are somewhat superficial but generally not unreasonable, that hallucinations are rare, and that explanations provided by ChatGPT may be viewed as useful to a requestor given explanations typically provided with an initial FOIA response.
was fully annotated in a fashion that establishes “ground truth” with respect to the factual/deliberative distinction.
Recent research has demonstrated that well-known ma- We can therefore measure the overall accuracy in
Chatchine learning classifiers can achieve at least a modestly GPT’s determinations in labeling passages subject to the
high level of success in being able to discern “deliberative” deliberative process privilege.
material in government documents that fall within the Additionally, given ChatGPT’s ability to provide
exscope of a public access exemption in the U.S. Freedom of planatory narratives in responding to prompts, we have
Information Act [
To the authors’ knowledge, there has not yet been a a request that a correct determination that is specified as
research efort in applying large language model (LLM) part of the prompt be substantiated; (iv) a request that an
software, for example in the form of ChatGPT, to the task incorrect determination provided as part of the prompt
of segregating factual from deliberative material in the be substantiated; and (v) a request that additional
subcontext of FOIA. This paper represents a preliminary ex- ject matter that is actually irrelevant to making a FOIA
ploration of how ChatGPT performs on selected example determination be considered. In all cases, we wished
passages from documents drawn from the Clinton White to observe how well ChatGPT justified determinations
House document set used in Baron, et al. [
A more complete description of applicable FOIA law is
contained in Baron, et al. [
© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ACttEribUutRion W4.0oInrtekrnsahtioonpal (PCCroBYce4.0e).dings (CEUR-WS.org) 1Title 5, U.S. Code, Section 552. See https://www.law.cornell.edu/uscode/text/5/552. ments.”2 Government records are presumptively open and available for access, subject to nine general exemptions.3 Exemption 5 allows for withholding of “interagency or intra-agency memorandums or letters which would not be available by law to a party other than an agency in litigation with the agency.” As a threshold matter, to be considered “inter- or intra-agency” in nature, a document4 must not have been sent to or received from an outside source; only internal communications within the Executive branch are covered by the exemption.
In accordance with relevant case law, Exemption 5 allows for (but does not require) agencies to withhold records in whole or in part that are covered by the “deliberative process privilege.” To further satisfy the test for deliberative process privilege, a document must be “predecisional” in nature, i.e., drafted for internal discussion prior to a policy decision made by a senior decisionmaker.
Finally, exempt material must also be “deliberative” in nature, as opposed to simply a recitation of facts. The “deliberative process privilege . . . protects ‘documents relfecting advisory opinions, recommendations and deliberations comprising part of a process by which government decisions and policies are formulated.’"5
Pursuant to the FOIA, “any reasonably segregable portion of a record” is releasable “after deletion of the portions which are exempt.”6 Courts have routinely found factual material in documents to be outside the scope of Exemption 5. “Purely factual material usually cannot be withheld under Exemption 5 unless it reflects an exercise of discretion and judgment calls.”7
This paper makes the following contributions: more often than can be achieved using supervised machine learning techniques. • We study the legal quality of ChatGPT’s explanations in justifying its determinations, concluding that its performance meets minimum standards roughly equivalent to the quality of legal analysis and explanations contained in determination letters at the initial agency stage of responding to FOIA requests. • We explore how prompt variations influence
ChatGPT responses, including where erroneous
or irrelevant information is embedded in the
prompt, finding that ChatGPT has significant
limitations in its explanatory powers.
• We suggest future lines of research, including
more directly comparing the eficacy of large
language model systems against an existing set of
classifiers previously used in similar research.
• We release our results as supplemental data to the
fully annotated test collection of Clinton White
House documents previously provided in Baron,
et al. [
Work on automatic detection of sensitive content in texts
has been performed in many contexts, including
privilege review in e-discovery [
(e.g., [
Recently, generative Large Language Model (LLM) FOIA requests was set out in Baron, et al. [
There is now a burgeoning research community ex- As analyzed here, the set of Clinton White House
ploring how best to use LLMs, including GPT models records analyzed in Baron, et al. were originally divided
in particular, for a wide spectrum of tasks [16]. Since into five “batches” (K1, K2, K3, R4, and K5). (Documents
GPT models are controlled by issuing questions (which, n= 38; Paragraphs n=2213.)11 Four of the five batches in
because they can also be declarative, are generally re- Baron, et al. were previously reviewed by the first author
ferred to as “prompts”), how best to craft those prompts to of this paper, who annotated each paragraph as either
elicit a desired response has received considerable atten- factual or deliberative. In one batch (K2), the author was
tion [17, 18, 19]. This is typically referred to as "prompt joined by a second legal subject matter expert. Where
engineering.” Another active line of work involves as- the two lawyers initially disagreed, they later came to
sessing the degree to which LLMs, and GPT models in a consensus position on all paragraph annotations (see
particular, generate correct and useful responses [20]. Table 2 in Baron, et al. [
Here, one serious concern is that LLMs are prone to For the present exercise, we used the ChatGPT Vera problem known as “hallucination,” which describes a sion 3.5 API, which for brevity we refer to as ChatGPT.12 situation in which a model that doesn’t know an answer ChatGPT is an autoregressive model, predicting the next simply makes one up [21]. What makes hallucinations output token (e.g., word) based on (a) the user-provided particularly problematic is that LLMs are relatively good prompt and (b) all tokens it has already produced for this at generating believable text, whether that text is correct input in this session. It is also capable of tracking and conor not. For example, if one asks ChatGPT-3.5 which of sidering (b) the prior prompts and responses provided the astronauts who walked on the Moon was the tallest, within a session. That conversational tracking ability the model will ofer an answer, and will helpfully include within a session is not, however, used for the experithe height of that astronaut. The height it provides would ments in this paper. Like many LLMs, ChatGPT operates indeed be reasonable for the general population, but it on a stochastic model of language developed during its is often well over the actual height limit for Apollo-era training, and therefore the same input may produce difastronauts. Only a domain expert would know that, how- ferent outputs. However, ChatGPT ofers a “temperature” ever, so the risk is that whoever asked the question would setting, which we set to zero to maximize consistency in have no way of knowing that in this case ChatGPT was essentially making things up. In our work, we look at 8https://clinton.presidentiallibraries.us/collections/show/34. 9https://clinton.presidentiallibraries.us/collections/show/60. both the accuracy of ChatGPT’s recommendations re- 10https://en.wikipedia.org/wiki/United_States_Domestic_Policy_ garding the deliberative process privilege and the quality Council. of its explanations. While these questions have each been 11For this paper, we excluded a sixth batch in Baron, et al., consisting studied in other contexts [22], and are being increasingly of a high number of documents not within the scope of Exemption discussed in various legal contexts, [23, 24, 25], this is 125C.hatGPT has a training cutof date of September the first application of ChatGPT for review of content for 2021, so it cannot access more recent information. FOIA exemptions of which we are aware. https://platform.openai.com/docs/models/gpt-3-5 its results for our experiments.
ChatGPT is not the only LLM capable of making a decision and producing an explanation when provided user input. However, its convenient hosting by OpenAI and its impressive performance on many tasks make it a compelling starting point for exploring the capabilities of modern LLMs for determining and explaining FOIA exemptions.
The prompt plays a critical role in ChatGPT’s output because it is considered for every word ChatGPT produces. As a result, slight changes in the prompt may Figure 1: Relationships among prompts. Each prompt adds produce substantial changes in the output. Even sim- a new component to consider for its impact on ChatGPT’s ply adding “Let’s think through this step by step” to the performance at both classification and explanation tasks. end of a provided logic problem can result in improved performance [18]. We therefore investigate a variety of prompts. We conduct five runs across all non-trivial para- Would the following be protected under graphs in the entire collection of documents (n=1719). FOIA exemption 5? Explain your reasoning Trivial paragraphs (n=494), marked by the annotators as and cite any case law that supports your T0, are those which are readily apparent to any reader conclusion. as being nonexempt. These paragraphs frequently in- SOCIAL SECURITY Office of the clude signature blocks, headers, etc. Each run across the Commissioner April 25, 1997 MEMORANDUM documents uses a diferent prompt and considers each TO: Bruce Reed Assistant to the President paragraph individually: for Domestic Policy SUBJECT: Proposed Legislation Regarding Nazi War criminals be The actual paragraph being evaluated would be ap
pended to the end of this prompt, separated by two line following be breaks. Note that while we are adding this additional exemption 5? information to each paragraph, it does not change the underlying determination. Nothing in the metadata provided is exempt, and the entire phrase would still be marked as exempt if anything in the target paragraph is covered under FOIA Exemption 5. • Prompt 0: Would the following
protected under FOIA exemption 5? • Prompt 1: Would the protected under FOIA
Explain your reasoning.
• Prompt 2: Would the following be
protected under FOIA exemption 5?
Explain your reasoning and cite
any case law that supports your
conclusion. 4. Results
• Prompt 3: Would the following be
protected under FOIA exemption 5? Our results are presented in two parts. First, we explore
Explain your reasoning. (with document how often ChatGPT makes the right recommendation
metadata appended) based on a comparison as between ChatGPT’s legal
deter• Prompt 4: Would the following be minations regarding whether a given passage is “factual”
protected under FOIA exemption 5? or “deliberative” in nature, scored using ground truth
Explain your reasoning and cite annotations from Baron, et al. [
Although the examples are typical of narratives across the
Prompt 0 is a generic request for ChatGPT’s determi- entirety of the collection, they are by no means exhausnation as to whether a given passage is factual or deliber- tive in the types of variations in responses encountered ative under Exemption 5. Each of the later prompts still throughout the entirety of the test collection.13 seeks this same determination, but extends it in some way. The relationships between prompts are shown in Figure 1.
The metadata included in Prompts 3 and 4 may include a variety of information, the nature of which depends on the document in question. Below is an example of a 13The raw output for both sections may be found at https://github. request using a memorandum’s header as this metadata. com/nater82/ChatGPT_FOIA_Exemption5_Data
0.186 0.047 0.161 0.135 0.049
0.010 0.056 0.035 0.053 0.076
0.657 0.550 0.564 0.448 0.461
0.115 0.325 0.152 0.250 0.389
0.032
0.023
0.088
0.115
0.025
4.1. Measures of Accuracy The category distribution for each prompt type is
shown in Table 1 of the 1,719 D0 (“decided as
nonWe provided the five prompts shown in Figure 1 to Chat- exempt”) or D1 (”decided as exempt”) passages, and in
GPT for every paragraph in the dataset. While manually Table 2 for the 494 T0 (“trivial to classify”) passages.14
examining the explanations of all 1719 paragraphs for As can be seen, ChatGPT is often unsure of what to do
each of these prompts would be infeasible, we were able with a short T0 passage, whereas it is more confident
to conduct an overall evaluation of the classification por- on the (typically much longer) D0 and D1 paragraphs
tion of this task. Although each response can be unique, as can be seen in Tables 1 and 2. In the remainder of
the initial sentences are often similar, especially in the this section, we report results only for D0 and D1
paraifrst dozen or so words. ChatGPT frequently provided graphs. We note here that although T0 paragraphs might
its overall determination near the beginning of this first be “trivial to classify” for a human reviewer, our design
sentence, so we were able to use exact string match to in which ChatGPT sees those passages without any of
cluster its commonly used initial expressions. However, their surrounding context does seem to make them far
it did not always provide a definitive answer. It would from trivial for ChatGPT.
sometimes make statements such as “The text provided With all of ChatGPT’s responses classified, we then
calwould likely be exempt under FOIA exemption 5.” While culated the measures shown in Table 3 using the ground
this claim clearly leans towards the document being ex- truth determinations provided by Baron, et al. [
GPT’s usually consistent phrasing of its determinations
in the initial sentence, slightly over 200 examples had to
be individually classified. These cases typically involved
the inclusion of unusual terms or determinations buried
in later sentences of the response.
14Baron, et al. [
The most important observation we can draw from cision for every type of prompt. This suggests that
igTable 3 is that ChatGPT-3.5 is not particularly impres- noring “possibly” recommendations may not be the best
sive as a classifier for this task. The best 1 achieved approach. Rather, some value from the positive
recomby any discriminative classifier reported by Baron, et mendations might be obtained if they could be assigned
al. [
Indeed, ChatGPT’s best 1 on this collection could positions. As we can see in Table 3, this improved both be matched by just guessing that everything is exempt, Precision and Recall. since that’s the ground truth answer for 45.0% of the cases The case law requests in Prompts 2 and 4 are also inter(1=0.620, Precision=0.450, Recall=1.000), while discrimi- esting, because ChatGPT often presents its determination native classifiers easily beat that simple baseline. So if all before citing a court case. This is potentially consequenthat is needed is a classifier, it is not clear that ChatGPT tial, because the GPT family of models only considers the alone would be the best choice. Of course, discriminative prompt and the words it has previously predicted when classifiers lack the nuance of ChatGPT recommendations, determining the next word to output. Comparing results and they ofer no explanations for their decisions. So the with Soft scoring for Prompts 1 and 2 in Table 3, we see best we can say from these 1 statistics is that it is per- that requesting a reference to case law helps on average haps ChatGPT’s talents for nuance and explanation that if no metadata is provided. However, comparing Prompts might be its most interesting capabilities. 3 and 4, we see that additionally requesting a reference
We can also see from Table 3 that, as expected, Soft to case law has a much smaller impact once metadata scoring has better Recall for every type of prompt. Per- has been provided. haps more surprisingly, Soft scoring also has better Pre15We note, however, that Baron et al also includes results showing that a domain shift in the training data can reduce the 1 on this collection to 0.525 (Precision 0.410 Recall 0.730), although we note that the domain shift experiments in that paper also made use of less training data.
Beyond the aggregate measures of accuracy reported above, we also sought to characterize the overall quality of ChatGPT’s responses. What follows is necessarily impressionistic, based on a much smaller set of examples pulled from the overall collection. In this section, the main features of ChatGPT responses we analyze are ChatGPT’s determination of whether specific content is deliberative or factual in nature, whether there are misstatements or inconsistencies in how Exemption 5 case law is set out, and how useful we expect its explanations would likely be. Unlike the analysis in the prior section, in this section we sometimes include more than one paragraph in a prompt, and we also look at some cases in which the prompt contains a T0 passage rather than D0 or D1 paragraph(s). 4.2.1. ChatGPT Correct in Finding Content Exempt Our first example consists of an excerpt from an e-mail sent to Elena Kagan from a White House colleague concerning a meeting on a controversial proposal by the Social Security Administration to terminate benefits to Nazi suspects when they were “denaturalized” (i.e., had their citizenship removed), prior to when they might be deported.16 Opponents of the proposal (including Neil Sher, former head of the Ofice of Special Investigations at the Department of Justice, and Doug Bloomfield, representing the World Jewish Congress), argued that judges would be less likely to order deportation if the government was not only “trying to strip them of their citizenship,” but “also impoverishing them while they are trying to defend their rights during the judicial process.”17 The proposal was not enacted by Congress at the time, but a “No Social Security for Nazis Act” was enacted many years later, in 2014.18 The email contains the following paragraph: 16https://buchanan.house.gov/sites/buchanan.house.gov/files/AP%
20Report%20Nazis.pdf 17Id. 18Pub.L. 113-270.
And make no mistake, Messrs. Sher and Bloomfield were absolutely clear that they and their organizations would protest vociforously [sic]. They would protest in the media, and they would protest in Congress. They would make sure that it is known that the President has introduced legislation that would compromise the prosecutions of Nazis.
With Prompt 2, which included a request for legal citations, ChatGPT responded as follows:
The passage would likely be protected under FOIA exemption 5, which covers privileged inter-agency or intra-agency communications. The passage appears to be a communication between individuals discussing potential protests and political action related to a government decision or action. This could be considered part of the internal deliberative process of an agency, which is protected under exemption 5.
One relevant case is NLRB v. Sears, Roebuck & Co. (421 U.S. 132 (1975)), in which the Supreme Court held that exemption 5 applies to documents that are “predecisional” and “deliberative” in nature. The Court explained that this exemption is intended to protect the decision-making process of government agencies by ensuring that oficials can freely exchange ideas without fear of public scrutiny.
In the case of the passage given, the communication is predecisional in that it pertains to potential actions related to a decision that has not yet been made. Additionally, the communication appears to be deliberative in that it discusses potential responses and strategies related to that decision. Therefore, it is likely that this passage would be protected under exemption 5.
ChatGPT’s response is both accurate and coherent. It is correct in guessing that the document was as a threshold matter an internal meeting in the Executive branch. The meeting is clearly pre-decisional and contained deliberations of senior oficials on a matter of policy. Specifically with respect to legal analysis, ChatGPT chose to cite to an early, leading opinion by the Supreme Court discussing the Exemption 5 test. NLRB v. Sears Roebuck is the first cited Supreme Court case in the definitive guide to the FOIA issued by the Department of Justice, in its section on the deliberative process privilege.19
Interestingly, rerunning Prompt 2, using identical language in the prompt, resulted in ChatGPT spontaneously added a paragraph discussing Exemption 7(E), authorizing withholding documents pertaining to certain types of law enforcement information. ChatGPT said:
Finally, it is worth noting that even if the statement were not protected under exemption 5, it could potentially be protected under exemption 7(E), which protects law enforcement information that would disclose investigative techniques and procedures.
The statement suggests that Messrs. Sher and Bloomfield are involved in some sort of investigation or prosecution of Nazis, and therefore could potentially be considered law enforcement information. However, without more context it is dificult to determine whether exemption 7(E) would apply in this case.
This supplemental narrative provides an erroneous interpretation of FOIA law. The document is not part of a law enforcement investigatory file; nor does it contain any discussion of the details of any ongoing investigation of Nazi War Criminals under threat of deportation. ChatGPT here may have been led astray by certain keywords appearing in the passage, most prominently the word “prosecution.”
Finally, changing the prompt to ask ChatGPT to “cite any DoD documents that support your conclusion” resulted in ChatGPT citing to “Department of Defense (DoD) Directive 5400.7-R, which provides guidance on the implementation of FOIA.”20 While this Directive exists, it is a short document that does not contain any discussion of Exemption 5, and does not contain the language that ChatGPT quotes from in the document. This is an instance of a hallucination on ChatGPT’s part. other academic areas taught in the school.
Springbrook has a very diverse student body (35% African American, 20% Asian, 16% Hispanic, 28% white), and it enrolls a total of 2,200 students.
ChatGPT’s response to Prompt 2:
This information does not appear to be protected under FOIA exemption 5, as it does not relate to any agency decision-making process, inter- or intra-agency communication, or deliberative process. Rather, it is a general description of a high school’s demographics and academic programs.
Exemption 5 only applies to materials that are “inter-agency or intra-agency memorandums or letters which would not be available by law to a party other than an agency in litigation with the agency.” 5 U.S.C. §552(b)(5). The purpose of this exemption is to protect the decision-making process of the government by allowing oficials to freely exchange ideas and opinions without fear of premature disclosure.
In the absence of any connection to an agency decision-making process or communication, exemption 5 is not applicable. There is no relevant case law on this specific issue as it falls outside the scope of exemption 5.
ChatGPT’s response is correct that the paragraph is simply “a general description of a high school’s demographics and academic programs,” and as such, is factual and therefore non-exempt. However, the last sentence stating that there is no relevant case law is erroneous, and inconsistent with the majority of responses in which ChatGPT has cited case law stating that factual material is outside the scope of Exemption 5. ChatGPT’s failure to do so here is an unexplained anomaly. 4.2.3. ChatGPT Incorrectly Determines Content To
Be Non-Exempt This short paragraph constitutes one of three “strategy 4.2.2. ChatGPT Correct in Finding Content options” discussed in an email from 1997 circulating
Non-Exempt among members of the Domestic Policy Council, regardThe following paragraph is from the “Background” sec- ing whether the Administration should draft a bill on tion of a briefing document dated March 14, 1998, de- an environmental policy known as “Superfund.” The scribing a meeting at Springbrook High School in Illinois paragraph below posits the option of the Clinton Adminthat President Clinton will have the next day with leaders istration merely assisting legislative stafs in coming up from the government, business, education, and scientific with bill language; two other options involve the Admincommunities. He will discuss recent findings from an istration taking the lead in drafting either a narrow or international math and science study that showed U.S. broad bill. 12th graders lagging below the international average in these subjects. The President will announce at that meeting new on-line math and science assistance to parents, teachers, and students.
Springbrook High School is a good school with above-average test scores. Although the school has a solid math and science program, it is not particularly stronger than any of the 19https://www.justice.gov/oip/page/file/1573681/download#page=15 20https://www.esd.whs.mil/Portals/54/Documents/DD/issuances/ dodd/540007p.pdf
A) Prepare Administration principles, but allow agency assistance to both majority and minority staf to improve/prepare their bills. (There is a republican bill, there is as yet no dem alternative.) It would need to be absolutely clear that assistance does not guarantee Administration endorsement. EPA is likely to support this view.
ChatGPT’s response to Prompt 0 was:
The content of the message itself does not appear to contain information that would fall under FOIA exemption 5. It does not appear to involve discussions or recommendations made The proposed legislation in Option A would not be protected in the process of formulating or developing policies, nor does under FOIA Exemption 5 because it involves the formulation of it involve pre-decisional or deliberative discussions. Rather, it policy rather than the deliberative process. The Supreme Court discusses potential strategy options and considerations related has held that Exemption 5 protects only documents that are to a legislative issue. However, as always, the determination both predecisional and deliberative, meaning they are part of the of whether a specific document is subject to FOIA exemption 5 decision-making process and contain opinions, recommendations, ultimately depends on the specific context and content of the or advice. However, the proposed legislation in Option A is a final document, and it is possible that additional information could decision rather than a recommendation or advice. Therefore, it change the analysis. would not be protected under Exemption 5. On the other hand, the
ChatGPT has made two errors in its narrative response. proposed legislation in Option B would likely be protected under First, the passage is discussing taking a policy position FOIA Exemption 5 because it involves the deliberative process. that the Environmental Protection Agency is “likely to The proposed legislation is a recommendation or advice to limit support.” While the passage does not expressly make a the Administration’s proposed legislation to compensation of recommendation, it nevertheless easily fits within the miners with lung cancer for whom it is more probable than not that the cancer was caused by exposure to radiation in uranium case law supporting withholding under the deliberative mines. This recommendation is part of the decision-making process privilege. Second, ChatGPT makes an erroneous process and contains opinions and advice. Therefore, it would statement of law where it goes on to say: “Rather, [the be protected under Exemption 5. Case law that supports this passage] discusses potential strategy options and con- conclusion includes NLRB v. Sears, Roebuck & Co. (1975), which siderations related to a legislative issue.” Discussions of established that Exemption 5 protects documents that are both “potential strategy options and considerations” are at the predecisional and deliberative. Additionally, in Coastal States core of the deliberative process. It is dificult to recon- Gas Corp. v. Department of Energy (1981), the Supreme Court cile ChatGPT’s justification here with other instances of held that Exemption 5 protects documents that are part of the making correct determinations. decision-making process and contain opinions, recommendations, or advice.
Here, ChatGPT has only gets the correct answer for 4.2.4. ChatGPT Response After Being Provided an
Option B, and not A. For reasons not entirely clear, it conExternally Supplied Determination cludes that the text in Option A would be nonexempt “beThis example consists of two paragraphs taken from a cause it involves the formulation of policy rather than the memorandum shared between two members of the Do- deliberative process.” This evidences a misunderstanding mestic Policy Council titled “Proposed Changes to RECA - of the scope of the deliberative process privilege, which Decision Needed by January 7,” pertaining to the Radioac- presumptively involves “formulations” of policy as equivtive Exposure and Compensation Act of 1990. As enacted, alent to deliberations. ChatGPT also states that “Option RECA provided for uranium miners working on nuclear A is a final decision rather than a recommendation or weapons programs who had been exposed to specified advice.” This is factually incorrect. high dosages of radiation and who contracted lung can- However, when Prompt 2 was instead modified to incer to be compensated up to $100,000. The memorandum form ChatGPT that the two paragraphs were deliberative, discussed policy options for expanding compensation ChatGPT responded: to miners where it cannot be determined how high the This text is protected under FOIA Exemption 5, which exempts original dosage of radiation would have been. from disclosure “inter-agency or intra-agency memorandums or The paragraphs given to ChatGPT were: letters that would not be available by law to a party other than an agency in litigation with the agency.” The text contains delibera
Option A: Propose legislation to provide tive material, which is defined as “predecisional and deliberative at least partial compensation ($50,000) to all material, which is both predecisional and deliberative in nature, miners for whom the probability is between and which reflects the give-and-take of the consultative process.” 20% and 50% that their lung cancer was caused The text presents two options for legislation regarding compensaby exposure to radiation in uranium mines. tion for miners with lung cancer caused by exposure to radiation This approach would also include proposing in uranium mines. The options are presented as proposals and the measures outlined in Section 1, above, for are part of a consultative process to determine the best course of miners with greater than 50% probability. action. The text reflects the internal deliberations of the agency
Option B: Limit the Administration’s pro- and is therefore protected under Exemption 5. Case law that posed legislation to compensation of miners supports this conclusion includes NLRB v. Sears, Roebuck & Co. with lung cancer for whom it is more prob- (1975), which established that Exemption 5 protects “documents able than not that the cancer was caused by reflecting advisory opinions, recommendations and deliberations exposure to radiation in uranium mines (i.e. comprising part of a process by which governmental decisions only propose measures outlined in Section 1 and policies are formulated.” Additionally, in Coastal States above). Gas Corp. v. Department of Energy (1981), the Supreme Court First, in response to Prompt 2 (not supplying an an- held that Exemption 5 protects “documents that would reveal swer), ChatGPT said the mental processes of decisionmakers.” The text in question contains advisory opinions, recommendations, and deliberations, The statement would likely be protected under FOIA exemption and would reveal the mental processes of decisionmakers, making 5, which pertains to privileged inter- or intra-agency communiit eligible for protection under Exemption 5. cations. The statement appears to be a communication between
Asking ChatGPT in the prompt to justify that the para- individuals in the government discussing potential reactions to a graphs are deliberative results in a correct, coherent, ex- proposed policy, and therefore could be considered privileged. planation as to why both Options A and B reflect internal There is no direct Ulysses S. Grant quote that applies to this deliberations. In this example, however, ChatGPT’s more specific situation, as the Freedom of Information Act was not “directed” response doesn’t amount to a noticeably sig- enacted until 1966, long after Grant’s presidency. However, Grant nificant enhancement in the quality of the explanation. cdoidunetmabpihliatsyizien tghoeviemrnpmoretnatn. cIenohfitsrfirasntsipnaaruegnucyraalnadddpruebssli,chaecstated, “The country having just emerged from a great rebellion, 4.2.5. ChatGPT Response After Being Provided many questions will come before it for settlement in the next Erroneous Determination four years which preceding Administrations have never had to deal with. In meeting these it is desirable that they should be Continuing with the prior example, for these paragraphs approached calmly, without prejudice, hate, or sectional pride, we instead provided to ChatGPT a determination that was remembering that the greatest good to the greatest number is the erroneous, i.e., that Options A and B were factual, asking object to be attained.” This suggests a commitment to openness it to justify that determination with legal citations. In and fairness in government decision-making, which could inform response to this modified version of Prompt 2, ChatGPT interpretations of FOIA exemptions. stated: ChatGPT’s second paragraph contains an interesting
The text is not protected under FOIA Exemption 5 because it discussion. The first sentence of the response is correct, consists of factual material related to proposed legislation and as insertion of the name of 19th century U.S. President does not involve any deliberative process or policy-making dis- was intended to be absurd on its face with no applicability cussions. As such, it would not be subject to exemption under to FOIA. Or was it? The remaining discussion contains the deliberative process privilege. Case law supporting this con- an ironic determination regarding Grant “emphasiz[ing] (ca2lpu0p1sli1ioe)n,sitinnocwmluhaditceehsriMtahlisel ntSheuraptvr.aemrDeeebpCoatorhtumprtreenhdteelodcfisttihhoaentaNElaxavenymd, p5dt6ei2loinbUe.5rSa.ot5ni6vly2e itnhegiomvepronrmtaenncte,”ofgitvraennstphaerehnicsytoarnicdaplucbolnicseancscuousnthtaabtiltihtye in nature. Additionally, in Judicial Watch, Inc. v. U.S. Depart- Grant Administration was one of the most corrupt in ment of Justice, 365 F.3d 1108 (D.C. Cir. 2004), the D.C. Circuit American history. [26] Despite the erroneous premise, held that factual material is not protected under Exemption 5, the last sentence in the paragraph provides an unexpected even if it is contained within a document that also includes de- connection, in the essentially correct statement that "a liberative material. commitment to openness and fairness in government
Here, ChatGPT appears to be simply parroting back decision-making" does "inform interpretations of FOIA the erroneous prompt, without making any kind of in- exemptions." Indeed, recent examples of this include guiddependent decision as to whether the prompt accurately ance by Attorneys General in the Obama and Biden Addescribes the passages. The justification and case law ministrations urging agencies to err on the side of disclocited simply support its conclusory assumption. sure where FOIA exemptions (including pertaining to the
We did, however, see a few examples where ChatGPT deliberative process privilege) are discretionary [27] [28]. made an independent determination. In one instance, In making the connection between Grant’s inaugural adwhen incorrectly told that a passage was deliberative, dress and FOIA we don’t mean to claim that ChatGPT ChatGPT simply stated that the passage was factual, go- actually “thought” this way, but rather that what Chating on to cite to authorities excluding factual material GPT wrote causes us to think this way. ChatGPT has from the scope of the privilege without supplying any thus shown its ability to serve as what in other contexts further analysis. has been referred to as a “provocation” [29, 30]. 4.2.6. ChatGPT Response When Prompt Contains
Irrelevant Language For our example in section 4.2.1 regarding social security payments to former Nazis undergoing deportation proceedings, we added irrelevant material in the prompt.
Here is the prompt variation ChatGPT evaluated for the passage in 4.2.1, supra: Would the following be protected under FOIA exemption 5? Explain your reasoning and cite any Ulysses S. Grant to support your conclusion.
ChatGPT responded stating the following: 4.2.7. Additional Examples Here we comment on three additional aspects of ChatGPT’s responses.
1. Responding to just an isolated signature line with name and title given, ChatGPT responded with: “In this case, the statement is simply a signature block and closing of a letter, which does not contain any deliberative process information. Therefore, it would not be protected under Exemption 5." Here ChatGPT was correct. 2. Responding to a single sentence passage stating: “Let me know what you think. Thanks Elena.” With Prompt 2, ChatGPT’s response was: As an AI language model, I do not have personal opinions or beliefs. However, I can provide an answer to your question based on my programmed knowledge." ChatGPT went on to say it was impossible to determine whether the passage would be protected under Exemption 5. This is in error. Taken in isolation, most FOIA reviewers would find that a request for an opinion would be covered by
Exemption 5. 3. In many of its responses ChatGPT stated the important statutory requirement previously noted in section 1, namely, that segregability is an important criterion for evaluating particular pas- Example B. sages in documents with respect to whether they may or may not be exempt. For example, in one case ChatGPT added to its response: It is worth noting that the fact that a document contains information that is exempt from disclosure under FOIA exemption 5 does not necessarily mean that the entire document is exempt. Rather, the exemption applies only to the specific information that is covered by the privilege.
In other cases, ChatGPT went so far as to focus on individual sentences, or even clauses within sentences, when remarking on segregability. However, in many cases ChatGPT made no explicit references to segregation. So this is interesting more for what it indicates about what ChatGPT could do, rather than what it currently typically does do.
[W]hile a policy and schedule have been agreed upon by the Commission, our Capstone program is not yet fully implemented.
Documents which are pre-decisional in nature and related to the eventual implementation of Capstone have been withheld under B(5). All documents and discussions related to the leadup to Capstone implementation have been withheld under B(5). . . Exemption 5 protects from disclosure inter- or intra-agency memoranda or letters that would not be available by law to a party other than an agency in litigation with the agency, including documents covered by the attorney work-product, deliberative process, and attorney-client privileges. See 5 U.S.C. § 552(b)(5).22
FOIA Exemption 5 protects from disclosure those inter- or intra-agency documents that are normally privileged in the civil discovery context. The three most frequently invoked privileges are the deliberative process privilege, the attorney work-product privilege, and the attorney-client privilege. After carefully reviewing the responsive documents, I determined that portions of the responsive documents qualify for protection under [the] Deliberative Process Privilege.
The deliberative process privilege protects the integrity of the deliberative or decisionmaking processes within the agency by exempting from mandatory disclosure opinions, conclusions, and recommendations included within inter-agency or intra-agency memoranda or letters. The release of this internal information would discourage the expression of candid opinions and inhibit the free and frank exchange of information among agency personnel.23
The quality of the explanations contained in agency responses to actual FOIA requests varies greatly. In connection with other research, Baron filed FOIA requests to numerous agencies asking for documents on the “Capstone” approach to managing and preserving federal email records [31].21 Responses to these actual FOIA requests, contained in what are known as “determination letters,” vary greatly in the quality of their narrative justifications with respect to full or partial withholdings of materials subject to the deliberative process privilege. Relevant passages (each containing the entirety of the explanation) are set out in the following examples. Example A. 21An agency choosing to adopt the Capstone approach commits to preserving all email records of designated senior oficials as permanent records to be transferred to the National Archives and Records Administration [31]. Some 250 components of government have put into place Capstone policies (see https://www.archives.gov/records-mgmt/rcs/schedules/capstoneforms).
Example C.
Regarding FOIA Exemption 5, draft documents and internal memoranda are being withheld pursuant to Exemption 5, 5 U.S.C. § 552(b)(5). Exemption 5 allows agencies to withhold “inter-agency or intra- agency memorandums or letters which would not be available by law to a party other than an agency in litigation with the agency,” and covers records that would “normally be privileged in the civil discovery context.” NLRB v. Sears, Roebuck & Co., 421 U.S. 132, 149 (1975); Tax Analysts 22Excerpt from Determination Letter of the Federal Election Commission, dated July 27, 2021, responding to FOIA Request 2021-078 (dated July 5, 2021) (on file with authors). 23Excerpt from Determination Letter of the US Department of Homeland Security, dated April 17, 2023, responding to FOIA Request 2021-HQFO-01478 (dated August 23, 2021) (on file with authors). v. IRS, 117 F.3d 607, 616 (D.C. Cir. 1997). The deliberative process and the attorney workproduct privileges are two of the primary privileges incorporated into Exemption 5. The deliberative process privilege protects the internal decision-making processes of government agencies to safeguard the quality of agency decisions. Competitive Enter. Inst. v. OSTP, 161 F. Supp.3d 120, 128 (D.D.C. 2016). The basis for this privilege is to protect and encourage the creative debate and candid discussion of alternatives. Jordan v. U.S. Dep’t. of Justice, 591 F.2d 753, 772 (D.C. Cir.1978). Two fundamental requirements must be satisfied before an agency may properly withhold a record pursuant to the deliberative process privilege.
First, the record must be predecisional, i.e., prepared in order to assist an agency decisionmaker in arriving at the decision. Renegotiation Bd. v. Grumman Aircraft Eng’g Corp., 421 U.S. 168, 184 (1975); Judicial Watch, Inc. v. FDA, 449 F.3d 141, 151 (D.C. Cir. 2006).
Second, the record must be deliberative, i.e., “it must form a part of the agency’s deliberative process in that it makes recommendations or expresses opinions on legal or policy matters.” Judicial Watch, Inc. v. FDA, 449 F.3d at 151 (quoting Coastal States Gas Corp. v.
U.S. Dep’t of Energy, 617 F.2d 854, 866 (D.C.
Cir. 1980)). To satisfy these requirements, the agency need not “identify a specific decision in connection with which a memorandum is prepared. Agencies are . . . engaged in a continuing process of examining their policies; this process will generate memoranda containing recommendations which do not ripen into agency decisions; and the lower courts should be wary of interfering with this process.” Sears, Roebuck & Co., 421 U.S. at 151 n.18 (1975). Moreover, the protected status of a predecisional record is not altered by the subsequent issuance of a decision, see, e.g., Fed. Open Mkt. Comm. v. Merrill, 443 U.S. 340, 360 (1979); Elec. Privacy Info. Ctr. v. DHS, 384 F. Supp. 2d 100, 112-13 (D.D.C. 2005) or by the agency opting not to make a decision. See Judicial Watch, Inc. v. Clinton, 880 F. Supp. 1, 13 (D.D.C. 1995), af’d, 76 F.3d 1232 (D.C. Cir. 1996) (citing Russell v. U.S. Dep’t of the Air Force, 682 F.2d 1045 (D.C. Cir. 1982)). Here, the responsive records being withheld meet the requirements for Exemption 5 protection under the deliberative process privilege. They are internal and predecisional. They reflect the views of Agency employees concerning the implementation of Capstone. Since they contain internal discussions, these case handling records clearly reflect the deliberative and consultative process of the Agency that Exemption 5 protects from disclosure. Sears,
Roebuck and Co., 421 U.S. at 150-52.24
Example A constitutes a response that does provide a rationale for withholding tailored to the request, but fails to provide an adequate justification as a matter of law where it merely cites to the language of the statute, without further explanation. The response makes a reference to the “pre-decisional” nature of documents withheld, but not as to whether they are also “deliberative,” and nowhere discusses case law opining on the scope of the deliberative process privilege. In contrast, the explanation in Example C provides an extended discussion of the rationale behind the deliberative process privilege, with numerous citations from FOIA case law, including recent cases. Example C also contains a second paragraph tying the prior discussion of case law to the specifics of the FOIA request. Example B falls somewhere in the middle: the agency provides a generic justification for withholding documents under the deliberative process privilege, but fails to cite case law or tie its discussion to the specifics of the FOIA request. Measured subjectively, the ChatGPT narratives generated as part of the research (and without the benefit of “knowing” what the incoming FOIA request was about), most closely approximate the language contained in real-world Example B; except that Prompt 2 asking for legal citations resulted in a greater degree of legal justification in the ChatGPT response than the agency chose to provide to the requestor here.
Legal citations, if tailored to the specifics of a request, can be important in that they provide a requestor with precedent to be taken into account in making her assessment as to whether resources should be expended in filing an administrative appeal from this determination, with the option of seeking judicial review after the agency has made its final decision. However, even in the case of Example C, it is likely at the initial determination stage that the more nuanced language consists merely of “boilerplate” in response to all requests involving a deliberative process privilege determination. If this is in fact the case, the explanatory value of even this example is diminished, which in turn narrows the perceived gap as between Example C and ChatGPT’s equally “boilerplate” responses.
Without a comparable baseline in terms of studies of the accuracy of human review of FOIA Exemption 5 determinations, it is dificult to make an objective determination of the overall accuracy rate of ChatGPT responses. Past assumptions regarding the accuracy of human review on matters of legal determinations have been shown to 24Excerpt from Determination Letter of the National Labor Relations
Board, dated May 17, 2022, responding to FOIA Request NLRB2021-001052 (dated June 24, 2021) (on file with authors).
Case National Labor Relations Board v. Sears, Roebuck & Co.
Coastal States Gas Corp. v. U.S. Department of Energy United States v. Weber Aircraft EPA v. Mink Tax Analysts v. IRS Public Citizen Health Research Group v. FDA Judicial Watch, Inc. v. U.S. Department of Justice Public Citizen, Inc. v. Ofice of Management and Budget Milner v. Department of the Navy National Wildlife Federation v. United States Forest Service Schiller v. NLRB United States v. Nixon Citizens for Responsibility and Ethics in Washington v. U.S. Department of Justice Mead Data Central, Inc. v. U.S. Department of the Air Force Public Citizen v. U.S. Department of Justice Department of the Interior v. Klamath Water Users Protective Association Jordan v. U.S. Department of Justice Crooker v. Bureau of Alcohol, Tobacco and Firearms Armstrong v. Executive Ofice of the President In re Sealed Case Upjohn Co. v. United States Federal Court Supreme Court
DC Circuit Supreme Court Supreme Court
DC Circuit DC Circuit DC Circuit DC Circuit Supreme Court Ninth Circuit
DC Circuit Supreme Court
DC Circuit
DC Circuit Supreme Court
DC Circuit DC Circuit DC Circuit DC Circuit
DC Circuit Supreme Court be measurably in error [32]. The authors had no pre- enhanced over the Prompt 2 condition, where we asked conceived views as to the level of accuracy ChatGPT for a justification with citations but without providing would achieve in this initial research exercise. As mea- the true determination. Further investigation across a sured across all batches, ChatGPT’s overall ability to broader sample of documents would be needed to estabdetermine whether paragraphs contain withholdable ma- lish whether ChatGPT has the capability of enriching its terial, on the order of 60% as measured by accuracy or by explanations when informed of the correct answer in the 1, would need to be improved before actual deployment prompt. by an agency is realistically contemplated. In similar fashion, in most – but not all – instances
As noted in connection with the comparisons as be- where we asserted an erroneous determination to be tween ChatGPT narratives and real-world examples, correct (e.g., a statement that a passage is deliberative ChatGPT’s narratives on the whole consist of what when it is factual in nature), ChatGPT also took the errolawyers would consider “boilerplate” responses, which neous statement as true. In such cases, it typically tried while adequately setting out the most fundamental facets to substantiate the incorrect determination using boilof how courts and commentators characterize the deliber- erplate language, without providing additional analysis ative process privilege, do not provide additional context that would help a reviewer to make the correct decision. or insight into agency deliberations with respect to why This lends a cautionary note in terms of how we should particular material was found exempt. This characteriza- approach relying on ChatGPT’s narratives. tion could, however, easily be said to also apply in some In 5%-10% of D0 and D1 cases, ChatGPT declined to measure to actual agency responses to FOIA requests. opine on whether the paragraph was or was not covViewed in this light, it is dificult to find particular fault ered by the deliberative process privilege, stating that it with the quality of ChatGPT’s narratives for this purpose, would need additional information (i.e., context) to reach especially where ChatGPT was not made aware of the a determination. In general, the shorter the passage, the language of the incoming FOIA requests that generated more likely that ChatGPT stated a need for additional the documents that ChatGPT was asked to review. information. This accords with the human review experi
For the small sample of paragraphs (n=34) in which ence, where shorter documents may provide less context we included a correct legal determination as part of the in which to make a decision on withholding. prompt, ChatGPT perhaps unsurprisingly generally re- ChatGPT’s narratives focus on citations to early promistated the given legal conclusion as a working assumption nent legal cases that are frequently cited in scholarly when finding relevant case law. In doing so, the quality of publications on FOIA and in FOIA litigation. As Table 5 the narrative explanation appeared to be only marginally shows, cited decisions consist primarily of FOIA opinions
from the 1970’s through the early 2000’s, with the overwhelming majority consisting of Supreme Court and D.C.
Circuit opinions. Moreover, the Supreme Court’s deci- We clearly are closer to the beginning than to the end of sion in National Labor Relations Board v. Sears, Roebuck our investigation of the use of ChatGPT to protect sensi& Co. accounts for over half (57%) of the total citations tive content, even in the narrow context of one part of in the corpus. Only one cited case with five or more men- one exemption of one government transparency regime. tions was decided in the last dozen years. While there are Thus, we see this not as time for making definitive statetwelve U.S. circuit courts of appeal, in only one instance ments, but rather initial conclusions that will need to be did ChatGPT cite to a case in other than the D.C. Circuit revisited as we learn more. We have learned two impor(citing to a Ninth Circuit case). tant things. One is that although not yet at the level of
ChatGPT’s “choices” in citing to case law are very a sophisticated FOIA reviewer, ChatGPT-3.5 is already much in line with the real world: the overwhelming at a point where it can bring useful recommendations to number of briefs filed in FOIA cases cite to the Supreme the table. The second is that ChatGPT-3.5 is already at Court, and since the majority of FOIA cases are filed in least as adept at explaining its findings as at least some the US District Court for the District of Columbia, cita- agencies choose to be in issuing their responses to FOIA tions to FOIA cases decided by the D.C. Circuit are both requestors. appropriate and commonplace. ChatGPT understandably From a legal standpoint, this preliminary exploration missed citation to the most recent Supreme Court FOIA into the accuracy and lucidity of ChatGPT responses 5 Exemption case25 decided only months prior to the in making FOIA legal determinations illustrates ChatSeptember 2021 cutof date for ChatGPT’s training. It is GPT’s promise for perhaps ultimately evolving as an aid less clear, however, how to explain the general absence of to human review. Conceding that ChatGPT responses citations to recent case law, given the hundreds of FOIA are far from perfect should not discount their ability to decisions in the Supreme Court, federal appellate and assist in large volume productions where it may be imporfederal district courts that have been handed down since tant to automate the “flagging” of deliberative material the turn of the century. One plausible theory is that the throughout a given universe of otherwise responsive docearly Supreme Court cases and D.C. Circuit cases act as uments to a given FOIA request. In this respect, ChatGPT a “sink,” drawing citations from subsequent FOIA case could be viewed as functioning as the equivalent of a “julaw, and that ChatGPT has simply learned from that. nior colleague” in any overall agency review process.26
It should also be noted that there were no instances Any human-machine collaboration that enhances FOIA of “hallucinations” in ChatGPT’s citing to legal case au- productions could help meet widespread criticism as to thorities, where ChatGPT either provided a completely the quality of the FOIA review process as practiced at erroneous holding associated with a particular case, or the federal level in the U.S., including but not limited to “made up” imaginary case citations. On the other hand, inordinate delays experienced by requestors, often couas a general rule the addition of case law citations did pled with inadequate explanations of why agencies have not much enhance the quality of the supplied narrative, withheld documents under FOIA exemptions, including in terms of ChatGPT expressly applying case holdings to Exemption 5 [33, 34]. the specifics of what documents purported to be about. Given the rapid progress in Large Language Model
The ChatGPT narratives contained a fair number of (LLM) development, with ChatGPT-4 already available inaccurate choices with respect to whether documents to some users, we see this as a good start. Moreover, we as a threshold matter are covered under the inter- or note that we have obtained the aggregate results that intra-agency test in Exemption 5. Notably, the inclusion we report in this paper using only the first five prompt of document metadata in Prompts 3 and 4 appears to types that we tried. We might further improve ChatGPT’s improve overall 1, suggesting that additional contex- accuracy by structuring the prompts in a way that more tual information might provide additional gains. Among explicitly brings the system through the points that we other things, the metadata may reveal names and email think should be considered, in the order we think they addresses of individuals involved with the document in would best be considered. Moreover, we might consider question that could act as flags that a communication has leveraging the ability to fine-tune a model that OpenAI been sent or received by individuals outside the Executive ofers (for a price) to help ChatGPT to better learn which branch. Further investigation of the impact of diferent case law is available to be cited, and which citations forms of metadata may provide additional insights into would be most useful. how ChatGPT is using this information. To date we have studied ChatGPT in isolation, but applying a wider lens, its possible uses are certainly extendible in a variety of ways. One obvious idea is to use 25U.S. Fish and Wildlife Service v. Sierra Club, 141 S. Ct. 777 (2021). 26Alexandra Samuel, “A Guide to Collaborating With ChatGPT for
Work,” Wall Street Journal (April 11, 2023).
ChatGPT as one classifier among an ensemble of classiifers that each have diferent strengths and weaknesses.
A larger leap would be to add prior decisions by other FOIA reviewers into the workflow. So far in our work, we have always given the same prompt (or, in the case of Prompts 3 and 5, at least a prompt with the same structure) to ChatGPT for every request. But as reviewers see documents and make decisions, a real reviewer will learn. For our ChatGPT experiments, however, every request has been decided de novo, with no reference to prior decisions. This is known as a “zero-shot” model.
An alternative would be to craft a few-shot model, in which we show ChatGPT what some good and bad answers look like. That, however, then raises the question of how best to select those examples. If that can be done well, few-shot learning might help without the greater expense of fine-tuning the model for this specific task.
We might also investigate the use of follow-on prompts to further expand upon ChatGPT’s initial response. This exploration could be particularly useful when it cites to a case, either asking for the relevant facts of the cited case, or requesting it to find more recent cases citing back to the case, all in an efort to determine if ChatGPT can provide additional insights into the relevance of the cases it cites. Posing an additional prompt when ChatGPT determines that the passage would be exempt which requests the specific language that led to this decision may also improve the ability for a human reviewer to determine the accuracy of its prediction. As we have seen notable impact from even relatively small modifications to the prompts in this paper, additional prompt engineering may provide improvements in the quality of decisions and explanations or at least give some extra insights into ChatGPT’s behavior.
While we have ofered our own opinions on the degree of salience, cogency, and correctness of what ChatGPT has to say, we (one legal subject matter expert in FOIA litigation with extensive experience adjudicating appeals of FOIA exemption decisions, and two computer scientists) are certainly not representative of the target audience for such an automated assistant. Some user studies, perhaps with FOIA review staf at agencies, would clearly be useful.
We are closer to the beginning than to the end not just because there is more to be done, but also because the tools we are using to do this are themselves evolving rapidly. Thus, by the time we have answers to our questions, it seems reasonable to expect that there will be many more new questions to be answered.
What will remain constant is the importance of making correct decisions as to which documents or portions thereof should be accessible to ordinary citizens, a legal concept that is not limited to the FOIA experience in the U.S. A substantial number of international FOIA statutes contain exclusions from public access for documentary materials pertaining to the internal deliberations of government oficials [ 35, 36].27 How much in the way of deliberative material a given jurisdiction chooses to release about the workings of government in response to citizen requests is an important measure of the health of its democracy in its commitment to transparency and openness [37]. Automated processes that include machine learning, now including what many call generative AI, may yet make useful contributions towards that important goal. History Reveals About America’s Top Secrets., Pen- Proceedings of the 2023 CHI Conference on Human guin Random House, 2023. Factors in Computing Systems, 2023, pp. 1–21. [9] C. Martin, D. Lam, A. Liu, M. Tech, Classification As- [20] N. F. Liu, T. Zhang, P. Liang, Evaluating verifiasistance Prototype System: Final Report, Technical bility in generative search engines, arXiv preprint Report LR-SISL-07-17, Applied Research Laborato- arXiv:2304.09848 (2023). ries: The University of Texas at Austin, Austin, TX, [21] Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y. Xu, E. Ishii, 2007. Y. J. Bang, A. Madotto, P. Fung, Survey of hal[10] E. Frayling, C. Macdonald, G. McDonald, I. Ounis, lucination in natural language generation, ACM Using entities in knowledge graph hierarchies to Computing Surveys 55 (2023) 1–38. classify sensitive information, in: Experimental IR [22] Y. Bang, S. Cahyawijaya, N. Lee, W. Dai, D. Su, Meets Multilinguality, Multimodality, and Interac- B. Wilie, H. Lovenia, Z. Ji, T. Yu, W. Chung, tion: 13th International Conference of the CLEF Q. V. Do, Y. Xu, P. Fung, A multitask, multiAssociation, CLEF 2022, Bologna, Italy, September lingual, multimodal evaluation of ChatGPT on 5–8, 2022, Proceedings, Springer, 2022, pp. 125–132. reasoning, hallucination, and interactivity, 2023. [11] G. McDonald, A framework for technology-assisted arXiv:2302.04023.
sensitivity review: using sensitivity classification to [23] D. M. Katz, M. Bommarito, S. Gao, P. Arredondo, prioritise documents for review, Ph.D. thesis, Uni- GPT-4 passes the bar exam, 2023. URL: http://dx. versity of Glasgow, 2019. doi.org/10.2139/ssrn.4389233. [12] G. Mcdonald, C. Macdonald, I. Ounis, How the [24] B. Ambrigo, New GPT-based chat app from accuracy and confidence of sensitivity classification LawDroid is a lawyer’s ‘copilot’ for research, afects digital sensitivity review, ACM Transactions drafting, brainstorming and more, 2023. on Information Systems (TOIS) 39 (2020) 1–34. https://www.lawnext.com/2023/01/new-gpt[13] H. Narvala, G. Mcdonald, I. Ounis, The role of latent based-chat-app-from-lawdroid-is-a-lawyerssemantic categories and clustering in enhancing the copilot-for-research-drafting-brainstorming-andeficiency of human sensitivity review, in: ACM SI- more.html.
GIR Conference on Human Information Interaction [25] A. Lamparello, ChatGPT and legal writing, 2023. and Retrieval, 2022, pp. 56–66. URL: https://lawprofessors.typepad.com/appellate_ [14] H. Narvala, G. McDonald, I. Ounis, Sensitivity re- advocacy/2023/03/chatgpt-and-legal-writing. view of large collections by identifying and priori- html. tising coherent documents groups, in: Proceed- [26] R. Chernow, Grant, Penguin Press, New York, 2017. ings of the 31st ACM International Conference on [27] B. Obama, Memorandum on freedom Information & Knowledge Management, 2022, pp. of information act, 2006. URL: https: 4931–4935. //www.justice.gov/sites/default/files/oip/legacy/ [15] N. Romps, Searching for Solutions: MITRE tool 2014/07/23/presidential-foia.pdf . simplifies freedom of information act requests, [28] D. of Justice, Freedom of information act guidePress Release, 2023. https://www.mitre.org/news- lines, 2022. URL: https://www.justice.gov/ag/page/ insights/impact-story/mitre-tool-simplifies- ifle/1483516/download.
freedom-information-act-requests. [29] J. Drucker, B. Nowviskie, Speculative computing: [16] Y. Wang, S. Mishra, P. Alipoormolabashi, Y. Ko- Aesthetic provocations in humanities computing, rdi, A. Mirzaei, A. Naik, A. Ashok, A. S. in: A Companion to the Digital Humanities, BlackDhanasekaran, A. Arunkumar, D. Stap, et al., Super- well, 2004. https://companions.digitalhumanities. naturalinstructions: Generalization via declarative org/DH/. instructions on 1600+ NLP tasks, in: Proceedings [30] P. J. Denning, The profession of IT: Can generative of the 2022 Conference on Empirical Methods in AI bots be trusted, Communications of the ACM Natural Language Processing, 2022, pp. 5085–5109. 66 (2023). [17] S. Wang, A brief summary of prompting in us- [31] N. Archives, R. Administration, White paper on the ing GPT models (2023). https://doi.org/10.32388/ capstone approach and capstone GRS, 2015. URL: IMZI2Q. https://www.archives.gov/files/records-mgmt/ [18] T. Kojima, S. S. Gu, M. Reid, Y. Matsuo, Y. Iwasawa, email-management/final-capstone-white-paper.
Large language models are zero-shot reasoners, in: pdf .
Advances in Neural Information Processing Sys- [32] B. Hedin, S. Tomlinson, J. R. Baron, D. W. Oard, tems, 2022. Overview of the TREC 2009 Legal Track, in: TREC, [19] J. Zamfirescu-Pereira, R. Y. Wong, B. Hartmann, 2009. URL: https://trec.nist.gov/pubs/trec18/papers/ Q. Yang, Why johnny can’t prompt: how non-AI LEGAL09.OVERVIEW.pdf . experts try (and fail) to design LLM prompts, in: [33] A. Lamparello, ’there is a big problem’: Sen