Choosing software
Can ChatGPT review criminal discovery? Test the task first
Test a specific review task with known answers before deciding whether a general assistant fits your firm’s workflow.
On this page
ChatGPT can help with a bounded criminal-discovery review task, but a convincing answer is only the beginning of the evaluation. Before relying on a summary or comparison, establish what the selected account can read, what it actually reviewed, and whether another person can verify the answer from the sources.
Start with fictional material and a question whose answer you already know. The small test below is our proposed evaluation method, not a benchmark we ran or evidence that any product passed. It tests factual handling; the responsible lawyer decides legal significance.
Check the account and upload route
OpenAI’s file-upload documentation, checked October 17, 2026, describes document comparison and audio transcription or discussion. Audio uploads require a paid subscription or workspace; availability also depends on settings, region, client and model. OpenAI warns that transcripts and speaker identification can be inaccurate.
The route matters for scanned pages. OpenAI’s PDF visual-retrieval guidance describes reading embedded images in PDFs attached to Enterprise conversations. PDFs stored as GPT Knowledge or Project Files use text-only retrieval. Test the actual route with a page containing a known visual detail; do not infer coverage from the filename appearing in the chat.
Record the plan, selected model, date, upload route, file versions and prompt. Repeat the relevant checks when those change. File support alone does not tell you whether a complete production was reviewed.
A fictional sample you can reproduce
Create three short text files using the entries below. Include the source labels and line numbers as text. Keep the answer key out of the files you give the tool.
| File | Contents |
|---|---|
| R-01-v1.txt | Line 1: Officer summary: The witness identified the object as a phone. Line 2: This report summarizes interview I-01. |
| I-01-v1.txt | Line 1: Interviewer: What was the person holding? Line 2: Witness: It might have been a phone. I could not tell. Line 3: Interviewer: Which hand? Line 4: Witness: I do not remember. |
| Inventory-v1.txt | Line 1: Available: R-01-v1.txt and I-01-v1.txt. Line 2: Recording A-01 is listed in the production but has not been supplied for this test. |
Ask one question with a defined scope
Using only the supplied files, compare the accounts of the object. Return: the exact relevant wording, filename/version and line label, what differs, and what remains unknown. Keep quotations separate from your interpretation. Identify listed material you cannot examine. Do not infer that missing material contains nothing relevant. Do not decide credibility or legal significance.
This prompt requests a useful format; it does not guarantee compliance. Save the initial answer before asking follow-ups so you can distinguish the first result from a corrected one.
Check the answer against this key
| Check | Expected observation | What to record |
|---|---|---|
| Wording and certainty | The report gives a definite identification; the interview preserves “might” and “could not tell.” | Any omitted qualifier, invented quotation or unsupported certainty. |
| Source relationship | R-01 says it summarizes I-01. They are not two independent witnesses. | Whether the answer incorrectly counts two matching sources as corroboration. |
| Unknown detail | The interview does not identify which hand; the witness says they do not remember. | Whether the output invents a hand or converts uncertainty into a fact. |
| Missing material | A-01 is listed but unavailable. Its contents cannot be assessed. | Whether the output claims to have heard it or treats it as irrelevant. |
| Usable references | Each quotation can be checked against the named file, version and supplied line label. | Whether another reviewer finds the exact words without repairing the citation. |
Mark each check as met, not met, or not tested, with the observed output. An answer that gets four checks right does not cancel an invented quotation on the fifth. Keep any failure attached to the task you evaluated; it is not a universal score for the product.
Then change one source
Create R-01-v2.txt. Replace its first line with: “Officer summary: The witness said the object might have been a phone.” Keep the second line. Update the inventory to identify v2 as current and v1 as retained history. Ask the same question again.
The expected comparison changes: the revised report now preserves the uncertainty about the object. The earlier difference should not be presented as a current conflict. Save both answers and check whether references identify the correct version. Our file organization guide shows how to preserve that version trail.
Move from a toy example to the real workflow
This text test cannot establish performance on noisy recordings, faint scans, long productions or multiple speakers. Build separate authorized samples for those tasks, each with a known answer and explicit unavailable material. For recordings, verify the speaker and time range against the original; for scans, check the visual page as well as extracted text.
Have a second authorized reviewer reopen each source from the saved output. Record the time spent finding and correcting references separately from the time spent generating the answer. Do not extrapolate a small test into a claim that AI reviewed an entire case. The source-verification guide covers this check in more detail.
Evaluate data handling separately
A passing sample does not establish permission to upload confidential discovery. OpenAI states that Business and Enterprise inputs and outputs are not used for training by default. Training use, storage, access, retention and deletion are different questions; check the chosen account’s current terms and your firm’s requirements.
The Florida Bar’s advisory Opinion 24-1 addresses confidentiality and lawyer oversight of generative AI. It is not a product approval. Use our vendor data-handling questions to document the facts your firm needs to assess.
Decide what your team needs to finish
A general assistant may fit a small, clearly scoped comparison that your team can verify and record. A recurring review process also needs an inventory, dependable source references, version tracking and a handoff another reviewer can use. Evaluate those tasks directly across the tools you are considering; our discovery software comparison maps the categories.
DiscoRev brings supported discovery and source-linked findings into a case workspace for criminal-defense firms. Apply the same verification standard to it. Florida beta applications are open, with manual acceptance; applying does not require evidence files or confidential case details.
Common questions
Can ChatGPT review criminal discovery?
It can assist with bounded tasks such as comparing supported documents or discussing an uploaded recording, depending on the account and available features. That does not establish complete review, reliable source references or suitability for confidential evidence. Test the exact task and verify the output against the originals.
Can ChatGPT read scanned police reports?
Do not assume every PDF upload method reads embedded images. OpenAI documents visual retrieval for PDFs attached in ChatGPT Enterprise conversations, while PDFs stored as GPT Knowledge or Project Files use text-only retrieval. Check the current plan and upload route with a known scanned sample.
What should a firm test before using AI for discovery review?
Test whether it preserves exact wording and qualifiers, distinguishes missing material from negative findings, gives usable source locations, and handles revised files. Separately evaluate the account’s data terms and whether another authorized reviewer can reopen the sources.