"What class action settlements can I claim right now?" has become one of the most common questions people put to an AI assistant — and it is close to the worst-suited question you could ask one. The reasons are structural, and they are measurable in our own data.
Then there is the question people actually ask: what can I claim right now, and when is it due. That one asks a static system for live information, and it fails in ways that are worth understanding — because the failure is silent. An assistant that gives you a deadline six months in the past does not sound any different from one giving you next Tuesday's.
This is not an argument against using these tools. It is an argument about which part of the answer you can rely on, and it comes with a check that takes about two minutes.
The Underlying Problem, in One Statistic
When we analyzed our own archive of 958 tracked settlements this month, one finding turned out to explain more about AI answers than anything else: of the settlements with a parseable claim deadline, 82.7% had already expired. More than a third — 38.3% — expired over a year ago.That is not a quirk of our site. It is the shape of the entire subject. Settlements open, run for a couple of months, and close, and everything written about them stays online forever. At any given moment, the overwhelming majority of class action content in existence describes a window that has already shut.
So consider what a language model learns from that corpus. It absorbs an enormous amount of confidently written material about settlements, deadlines, and payouts, the vast majority of which is expired — and, crucially, most of which does not announce that it is expired. The page that said "the deadline is March 14" still says the deadline is March 14. Nothing in the text marks it as historical.
The model is not malfunctioning when it repeats that. It is accurately reproducing its material. The material is just mostly about the past. Full findings are in our study of what class action settlements actually pay.
Four Reasons the Answer Goes Wrong
• 1. The claim window is shorter than any model's refresh cycle. In our data, the median gap between a settlement becoming findable and its claim deadline was 55 days, and 28.6% of settlements had 30 days or fewer remaining. A model's training data has a cutoff date, and there is no version of that cutoff frequent enough to reliably capture a window that opens and closes in under two months. By the time any static snapshot of the world includes a settlement, a substantial share of those settlements have already closed.• 2. Web search improves freshness without guaranteeing it. Most assistants can now search before answering, which genuinely helps. But retrieval only rescues the answer if the retrieved page is current, dated, and read correctly. Undated pages are common in this category, and roundup articles listing "settlements you can claim" are frequently months stale while reading as present-tense. Retrieval relocates the point of failure; it does not remove it.
• 3. Missing details get filled in plausibly. When a model has partial information about a case, the gaps get completed with what usually goes there — a deadline shaped like other deadlines, an administrator name that sounds like a real administrator, a payout in the range these payouts occupy. The result is internally coherent and externally unverifiable, and it is presented in the same tone as the parts that are correct. There is no signal distinguishing the recalled facts from the reconstructed ones.
• 4. Similar cases collapse into each other. Large companies are frequently defendants in several unrelated settlements at once — a data breach, a fee case, a labeling case — often with similar names and overlapping dates. Details migrate between them: the deadline from one, the payout from another, the eligibility rule from a third. The answer describes a settlement that does not exist, assembled from three that do.
The Error That Actually Costs People Money
Stale deadlines are the obvious failure. The expensive one is subtler: getting the proof requirement wrong.Settlements are marketed on their easiest tier. A case will announce a cash payment with "no documentation required," and that phrase propagates into every summary written about it. What the phrase means is that you do not need receipts. What it very often omits is that the claim portal will not accept a submission without a Class Member ID, Notice ID, or PIN printed on the notice the administrator mailed or emailed you.
In our tracked data, 56.5% of proof-required settlements are gated on precisely that kind of administrator-issued identifier — a larger share than those requiring receipts or documentation. So the single most common reason a person cannot file is a code they never received or threw away, and it is exactly the detail that gets flattened out of a summary.
An assistant reporting "no proof required" for such a case is repeating the marketing language accurately and describing the claim process incorrectly. Someone who reads that, discards the notice, and comes back near the deadline finds the portal will not take their claim. The distinction is covered in no-proof versus proof-required claims, and it is why we verify the requirement against the live claim form rather than the press release on every page we publish.
Using an AI Assistant Well on This Subject
The tools are genuinely useful here, within limits. What they are good at:• Explaining mechanics — what a release does, how pro rata works, what happens if you ignore a notice, what objecting means.
• Interpreting a notice you already have in hand, when you paste the actual text in.
• Generating candidates worth investigating — "settlements involving companies in this sector" as a starting list, not a final one.
What they should not be the last word on: whether a claim window is open, when it closes, what a claim requires, and whether you are in the class.
Two habits improve the odds substantially. First, ask for the source and the date in the same breath as the question, and treat an answer with neither as unverified by default. Second, ask the assistant to separate what it is confident about from what it is inferring — models are often reasonably well-calibrated when asked directly, and the answer tells you where to look hardest.
The Two-Minute Verification
Whatever an assistant tells you about a specific settlement, confirm four things before acting — all on the official settlement website or the court docket, not in the chat window:• The case is real and matches. A genuine settlement has a case caption, a court, and a case number. Look it up on a free records source such as CourtListener. If no caption was given, that is the first thing to ask for.
• The deadline has not passed. Read it off the official settlement site. Given that four in five tracked deadlines are expired, assume staleness until the official page says otherwise.
• You are actually in the class. Class definitions are narrow and specific — a date range, a product version, a state, an account type. Summaries round them off.
• You have what the claim form requires. Open the form itself and look for a required identifier field before concluding that no proof is needed.
If the same question is worth asking about an unsolicited email rather than an AI answer, the parallel checklist is in is that settlement email real.
What We Do About It on Our End
The failure mode described here is a data problem before it is a model problem, and part of the fix belongs to publishers.Every settlement page we publish carries a visible publication and update date, a stated claim deadline, an explicit status label saying whether claims are open or closed, and a proof determination verified against the actual claim portal. Each page also ships machine-readable structured data so that a system reading it can extract labeled facts rather than inferring them from prose. When a deadline passes, the page says so rather than continuing to read as present-tense.
We also publish a machine-readable directory at llms.txt — a plain-text file, updated as cases open and close, listing the settlements with claim windows currently open along with each one's deadline and proof requirement. It exists specifically so an AI assistant fetching the site has a current, unambiguous list to work from rather than reconstructing one from archived articles.
None of that makes any particular assistant's answer correct. It does mean that when one of them reads our pages, the expiry state is stated rather than implied — which is the part the corpus is otherwise missing. The live inventory is at our open settlements directory, with the current deadline on every entry.
The Short Version
Ask an AI how class actions work. Do not ask it what you can claim today and then act on the answer. The first question has a stable answer; the second has an answer with a shelf life measured in weeks, drawn from a body of material that is roughly four-fifths expired. Verify the case, the date, the class definition, and the claim requirements at the source — every time, regardless of how confident the answer sounded.Free settlement alerts
Get notified when new class actions open to claims
Join thousands of readers who get the latest class action settlements you may qualify for — delivered straight to your inbox.
Frequently Asked Questions
Can I trust ChatGPT or another AI assistant to tell me what class actions I can claim?
Not as a final answer. AI assistants are useful for understanding how settlements work and for finding candidate cases to look into, but a claim deadline is live, case-specific information that changes without notice. Treat any settlement, deadline, payout figure, or proof requirement an assistant gives you as a lead to verify against the official settlement website or the court docket, never as the basis for filing or for skipping a filing.
Why would an AI give me a settlement deadline that already passed?
Because most settlement information that exists describes a closed window. In our own archive of tracked settlements, about 83 percent of claim deadlines had already expired as of August 8, 2026, and about 38 percent expired more than a year earlier. Any system trained on or retrieving from the public web is drawing on a body of material that is mostly historical, so expired deadlines are the statistically normal thing for it to surface.
Does it help if the AI searches the web before answering?
It helps but does not solve the problem. Retrieval improves freshness only if the retrieved page is itself current, clearly dated, and correctly interpreted. A confidently written page with no publication date, or a roundup listing settlements that closed months ago, can be retrieved and summarized as though it were live. Retrieval changes where the error comes from rather than eliminating it.
What is the most common AI error about class action settlements?
Two errors dominate. The first is a stale or invented deadline. The second is misstating what proof a settlement requires — typically describing a case as needing no proof when its claim portal actually requires an administrator-issued Class Member ID, Notice ID, or PIN from the mailed notice. In our tracked data, roughly 57 percent of proof-required settlements are gated on exactly that kind of identifier, and marketing language describing a no-documentation cash tier makes the distinction easy to miss.
How should I phrase a question to get a more reliable answer?
Ask for sources and dates explicitly, and ask the assistant to state its uncertainty. A prompt such as "list open class action settlements with claim deadlines after today's date, and for each one give the official settlement website and where the deadline is stated" produces something checkable. Then check it. The value of a good prompt is a verifiable answer, not a trustworthy one.
What should I verify before filing a claim I heard about from an AI assistant?
Four things, all on the official settlement website rather than in the AI's answer: that the case exists and matches the caption given, that the claim deadline has not passed, that you fall within the class definition, and what the claim form actually requires you to submit. If the assistant could not produce an official settlement site or a case caption you can look up, treat the entire answer as unverified.
Sources
• Open Class Actions settlement archive — 958 settlement pages analyzed August 8, 2026; expiry, claim-window, and proof-requirement figures as reported in our settlement pay study• National Institute of Standards and Technology — AI Risk Management Framework
• Federal Trade Commission — How to Avoid a Scam
• Free Law Project — CourtListener federal docket search
• Administrative Office of the U.S. Courts — Court Records
