By Jean-Philippe Beaudet
On July 29, The Digital Chamber’s AI and Quantum Working Group submitted comments to the Federal Trade Commission on its proposed policy statement addressing artificial intelligence accuracy.
The Proposed Statement treats bias mitigation as potentially deceptive. This premise does not match how AI systems are built, tested, or deployed. Model outputs, instead, reflect the models’ training data and choices of their human designers.
Our members already run model validation, bias testing, and remediation to meet federal and state law, contractual terms, and enterprise risk standards. Ensuring facial recognition systems supplied to government agencies can recognize all phenotypical indicators equally (they work on both black and white faces) requires fine-tuning models based on expected population demographics, for instance. A policy that casts that work as suspect would put them in conflict with obligations other agencies already impose.
We advised the Commission that:
- There is no universal neutral baseline. Every model output reflects the data and the choices that produced it. Across platforms, countries, and over time, we can see that untreated models, trained on historically biased data, reproduce those biases. The empirical record on lending, hiring, healthcare, and pricing models supports this.
- Treating mitigation as deception reverses the logic of Section 5. Consumers expecting neutral, objective outputs would not be served by FTC actions that mandate inaccurate model outputs. Section 5 protects consumers from deceptive practices, unfair competition, and operations that could violate their civil rights protections – like reducing the quality of their response based on their gender. There is a greater risk in presenting an unmitigated system as an objective score than in treating these outputs.
- For example, researchers in the world-renowned Nature journal recently found that, “when generating and evaluating resumes, [a leading LLM] assumes that women are younger and less experienced, rating older male applicants as of higher quality.”
- As AI is used both to prepare job documents and to review them, this tendency offers a prime example of the risks of unmitigated bias in AI workforce applications.
- The Proposed Statement cannot be read apart from the rollback of disparate-impact liability. Disparate impact is a legal concept that refers to a policy or practice that looks fair and nominally treats everyone equally but harms a protected group more than others in practice. In a disparate impact claim, you do not need to prove intention; you only need to show that the final result is unfair. Because AI has neither personhood nor intention, disparate impact treatment is often the only viable route for challenging algorithmic discrimination.
- Section 5 should reach material misrepresentations, not mitigation itself. The Commission can pursue firms that misrepresent what their systems do without treating responsible testing as presumptively deceptive.
- A reasonable federal floor beats broad preemption. TDC supports harmonization between a coherent national regulatory floor that balances innovation with consumer protections and coordinated state regulation. Industry concerns about regulatory fragmentation can and should be addressed through tiered, multi-state alignment on specific legislative remedies.
Read the full comment letter here.
If you have any questions, please reach out to policy@digitalchamber.org.