FOI BLOG

A chat about ChatGPT and FOIA


AI is useful for FOIA requests, but it still needs a human’s touch

UPDATE: Attorney Andrew Bodoh has created an AI protocol — instructions for your preferred AI tool for where to look and what NOT to do. Read his explanation here, and follow the links on his page for downloading and using the protocol.


I had a chat with ChatGPT today, and I have to say that I think I liked it better when AI was the bastardized start of the Spanish exclamation of woe: Ai yi yi. Because AI is bastardizing the FOIA request process.

I’m usually pretty hard on public bodies and officials in this space, but today I’m taking aim at requesters, even though it’s not totally their fault.

VCOG’s Newsletter is a reader-supported publication. To support our work, consider becoming a paid subscriber.

FOIA doesn’t require any particular formatting or wording, but a lot of people don’t know that. And FOIA can seem intimidating to some people. They don’t want to sound dumb when they make their request. They want their request to have gravitas. Many of them think they have but one bite at the apple to get everything they can possibly think of in an attempt to understand a particular decision or issue.

So it makes perfect sense for them — especially the uninitiated — to turn to AI to help them craft their FOIA response.

But I’m imploring them not to. Or, rather, I’m imploring them to pause and review what AI generates before copying and pasting it into their FOIA request and pressing “send.”

Chatting with a chatbot

We know that AI generates “slop” video and images. Well, it also generates “slop” FOIA requests that sound tough and official and legalistic, but that are, instead, complicated, overbroad and confusing.

Here’s what ChatGPT produced when I asked it to write a FOIA request to learn about a special use permit for a warehouse in my Virginia community. Here’s some of what it said (yes, it was actually much longer than what’s here):

Dear [FOIA Officer / Zoning Administrator / Clerk]:

Pursuant to the Virginia Freedom of Information Act, I request access to records concerning the application, review, and approval of the special use permit for the warehouse proposed or located at [property address / parcel number / project name].

…

Please provide the following:

  1. The special use permit application, including all attachments, site plans, architectural plans, traffic studies, environmental studies, and other materials submitted by the applicant.
  2. All staff reports, staff analyses, recommendations, and memoranda prepared in connection with the application.
  3. All correspondence and communications between the applicant, the applicant’s representatives, and [locality name] officials or employees concerning the application, including communications about revisions, conditions, concerns, or recommendations.
  4. All records reflecting the review of the application by county/city departments or agencies, including planning, zoning, transportation, public works, fire, emergency services, environmental, and other reviewing bodies, if applicable.
  5. All meeting agendas, packets, minutes, recordings, presentations, and public comments concerning the application or permit, including any public hearing.
  6. All records reflecting the basis for the decision to approve the permit, including findings, recommendations, decision memoranda, and any analysis of whether the application met the applicable zoning ordinance or comprehensive plan.
  7. The final special use permit and all conditions, proffers, agreements, or other requirements attached to its approval, including any subsequent amendments.
  8. All records concerning appeals, reconsideration, or challenges to the permit, if any.

Please include records in electronic format where available. If any responsive records are already publicly available online, please provide the specific links.

If any portion of this request is denied, please identify the specific statutory exemption authorizing the withholding and provide the nonexempt portions of the records.

If the request is expected to involve fees, please provide an estimate before incurring costs. If you believe the request is too broad, please contact me so I can narrow it.

Some of this isn’t necessarily slop, and I think it’s because ChatGPT knows who I am. Through my VCOG email account, I have asked questions and asked for output on a range of Virginia FOIA-specific issues. I’ve been training it on case law or statutory language it gets wrong. It knows what VCOG is and does. Spooky at times; helpful at others.

I call this slop because, as noted above, it is far too complicated, overbroad and confusing, and I told ChatGPT as much.

This FOIA request is written like a request for the production of documents during litigation. It’s complicated and long, and it includes terms and conditions that leave a lot to the interpretation of the government. In many ways, it is asking the FOIA officer to conduct an investigation. The FOIA request should be “reasonably specific,” but using terms like “other materials,” “in connection with,” “concerns,” “if applicable,” “reflecting the basis for,” are potentially limitless.

In true AI chatbot fashion, ChatGPT told me I was right (I should have asked it something that would have prompted it to tell me how smart and funny I am and what a good dancer I am). Then it said:

A good Virginia FOIA request should identify records by subject, custodian, date range, or type, rather than asking the public body to figure out what records might answer the requester’s underlying question.

For the warehouse example, I would make it much more like this:

I request the following records concerning the special use permit application for the warehouse at [address/project name]:

  1. The special use permit application submitted by [applicant], including the site plan submitted with the application.
  2. Staff reports prepared for the Planning Commission and/or Board of Supervisors concerning the application.
  3. Written comments submitted by members of the public concerning the application.
  4. Written comments submitted by government agencies concerning the application.
  5. The agendas and meeting minutes for [Planning Commission/Board] meetings at which the application was considered.
  6. The final resolution or other document approving the special use permit.
  7. The conditions imposed on the special use permit, including any proffers or development agreements incorporated into the approval.
  8. Revised versions of the site plan submitted between [date] and [date].

That’s much more FOIA-like: “Give me these identifiable records”, rather than “Give me everything that might explain this.”

The distinction is especially important with Virginia FOIA because the requester is entitled to existing public records; the FOIA officer isn’t required to conduct research, answer questions, create a record, or determine which universe of documents might be responsive to an amorphous inquiry.

I told ChatGPT that I thought this was better, but wasn’t it possible to condense or combine these eight categories of records? Again, it told me I was right (next time I’ll ask it to tell me how delicious my Monday Funday Casserole Delight is). And here’s what it came up with:

I’d group them according to the stage and function of the records, rather than trying to enumerate every possible document type:

  1. The application and revisions
    The application, site plans, and subsequent versions.
  2. The government’s review
    Staff reports and written comments from government agencies.
  3. The public process
    Public comments, agendas, minutes, and recordings of meetings at which the application was considered.
  4. The decision and its terms
    The approval document and the conditions, proffers, or agreements attached to it.

I’ll leave it to the FOIA officers reading this to tell me if and how this version could be made better (add your comment below), but I do think ChatGPT was getting to the right idea that a request should:

  • be reasonably specific;
  • include keywords, date ranges or people, when possible; 
  • ask for records, not answers; and
  • not ask the FOIA officer to make judgment calls about relationships between people or actions, much less conduct an investigation.

The advice I always give on top of those points is that the request doesn’t have to be crafted to capture absolutely every possible record. It makes much more sense to go in narrow, like a core sample in soil testing. Ask for a narrow range of keywords, people or dates. See what you get back. Then follow up with additional requests that are more targeted or strategic.

This isn’t You v. Them

It’s important to remember that the person who will fill this request is very likely not a lawyer and your FOIA request is not part of litigation. Neither requester nor receiver needs to treat the request process the same as if this were the discovery process in a court case.

To wit: I once saw a request that asked for five categories of records with a total of 25 subcategories; a three-month date range; literally hundreds of people; instructions the requester expected for format, search and processing; definitions of terms; an expectation that the request be filled in five working days; and a willingness to pay up to $100 for the request. The subcategories were rife with phrases like “concerning any condition, expectation, understanding, or consequence, or discussed in relation” to this, that or the other.

A FOIA officer, even a seasoned one who might even be a lawyer, is not prepared to handle this under the rubric of FOIA. They do not have the search tools; they are not necessarily close enough to the situation to know who, what, when or why people had understandings or discussed matters; they do not have the capability to ascertain what people they have no control over said or did.

Also, this request is not their only request, yet they are supposed to fill it within 5 days and for under $100?

I don’t know if AI was tied to this particular request, but that’s the direction some of these AI-generated requests I’m hearing about are headed. As with anything AI generates, good requesters will take the time to think about the scope of the output and will remember that FOIA offices are often understaffed and underresourced.1

AI promises to make the FOIA process easier. Users can get ideas for records to ask for and can eliminate the guesswork and the occasionally scary aspects of making a FOIA request. On the other side, AI promises to make the search, retrieval, review and redaction process that government uses more efficient. 

We’re not there yet, though, and until we are, let’s remember to be human.

1

Do I wish they were better staffed and funded? Absolutely! But I also have to deal with reality.