A consumer complaint succeeds partly through facts and partly through presentation. A study of 1,134,512 narratives submitted to the US Consumer Financial Protection Bureau between 2015 and 2024 found a sharp rise in language-model-assisted writing after ChatGPT’s release. Its instrumental-variable analysis estimated that LLM use increased the probability of favourable relief by 6.9 percentage points, with a 95% confidence interval from 4.9 to 8.9 points.[1][3]
The mechanism is unusually practical. Three controlled online experiments involving 1,010 US participants found that AI could improve complaint presentation without changing the underlying factual content. The technology did not create a better entitlement; it helped the consumer articulate the existing case more effectively.[1][3]
What the financial-complaint research measured
| Measure | Value | Context |
|---|---|---|
| Complaint corpus | 1,134,512 narratives | CFPB submissions from 2015 through 2024.Source 1 |
| Estimated relief effect | +6.9 percentage points | Instrumental-variable estimate; 95% confidence interval 4.9 to 8.9 points.Source 1Source 3 |
| Controlled experiments | Three | Designed to test how LLM assistance changes presentation.Source 1 |
| Experimental participants | 1,010 US adults | Total across the three online experiments.Source 1Source 3 |
| Mechanism | Improved presentation without altered facts | The experiments support clarity and advocacy quality as the pathway.Source 1Source 2 |
| Adoption pattern | Negative selection | Consumers otherwise prone to adverse outcomes were more likely to use LLM assistance.Source 1 |
The assistant can change the effective accessibility of a remedy
Complaint systems formally allow people to state what happened. In practice, they reward chronological clarity, relevant detail, explicit requests and language that makes the disputed action easy to evaluate. Consumers differ in confidence, English proficiency, time, document organisation and familiarity with financial institutions. An AI editor can reduce those differences without changing the governing rule.
That is why negative selection is significant. The people most likely to adopt LLM help were not simply the strongest writers seeking a small advantage. The study finds evidence that people otherwise prone to adverse outcomes were more likely to use it. A well-designed service may therefore operate as an equaliser rather than another premium convenience for already advantaged complainants.[1][2]
A high-standard complaint workflow
| Stage | Useful AI contribution | Evidence that must remain human-controlled |
|---|---|---|
| Chronology | Order dates, contacts, payments and promised actions | Original statements, transaction records and correspondence. |
| Issue classification | Identify the product, dispute type and requested remedy | The consumer confirms the category and what actually happened. |
| Drafting | Write a concise factual narrative and explicit resolution request | No invented facts, emotional claims or legal certainty.Source 1 |
| Evidence check | Link every important sentence to a document or memory note | Missing support stays marked as missing rather than being smoothed over. |
| Submission | Fit the institution’s fields and preserve a copy | The consumer approves the final text and recipient. |
| Follow-up | Summarise the response and deadlines | The system does not accept an offer or waive rights autonomously. |
Better writing should not become a new admissions test
The finding creates a policy dilemma. If clearer complaints receive better outcomes, making high-quality writing widely available is beneficial. It also reveals that the underlying system may be too sensitive to rhetorical quality. A valid claim should not fail because the consumer lacks the language of professional advocacy.
Institutions can respond in two directions. They can help complainants structure facts before submission, and they can train reviewers to evaluate the substance rather than fluency. An organisation could offer its own neutral complaint assistant, but it would need strong controls against steering users away from costly remedies or reframing disputed facts in the institution’s favour.
Keep the claim and the prose separate
- Create a private fact ledger before drafting: dates, amounts, people, communications and requested remedy.
- Require the model to label any inference, legal interpretation or missing evidence.
- Generate a plain version and a concise formal version from the same accepted facts.
- Check every number and quotation against the original record.
- Preserve the submitted version, attachments, receipt and response deadlines.
- Use AI to analyse a reply, but keep settlement, admission and waiver decisions with the consumer or adviser.
This is one of the clearest examples of AI increasing practical agency without taking the underlying decision. The consumer owns the facts and requested remedy. The model helps organise and communicate them. When that assistance is traceable and widely accessible, it can make a formal right easier to exercise—while reminding regulators that outcomes should not depend excessively on who can write like a professional advocate.
Sources & further reading
- The adoption and efficacy of large language models in US consumer financial complaints Nature Human Behaviour · 18 Feb 2026 · Accessed 11 Oct 2026
- Large language models have the potential to level the playing field in consumer financial complaints Nature Human Behaviour · 4 Mar 2026 · Accessed 11 Oct 2026
- The adoption and efficacy of large language models in US consumer financial complaints PubMed · Accessed 11 Oct 2026