Sen. Cruz Seeks Answers on FTC's Lawless Attempt at Regulating AI

Letter

Dear Chairwoman Khan:

I am writing to request information about the stance of the Federal Trade Commission
("FTC" or "Commission") on the regulation of artificial intelligence ("AI"). Your public
comments, as well as comments made to this Committee by senior FTC staff,1 suggest the FTC intends to play a role in aggressively policing AI despite receiving no explicit statutory
authorization to do so from Congress. As further evidence of the FTC's intent, on July 13, 2023,
a leaked Civil Investigative Demand ("CID") sent by the FTC to OpenAI--the California-based
company best known for its development of ChatGPT--shows the FTC is pursuing AI regulation
under legal theories that exceed the agency's statutory authority and would entail regulation of
constitutionally protected speech.2

Like many computer applications, AI is a productivity tool that is useless without human
guidance. In fact, ChatGPT assisted in drafting this letter. But AI computer code, apart from its
use by a consumer, has no inherent ability to violate the Civil Rights Act or Section 5 of the FTC
Act as your May 3rd op-ed in the New York Times, titled "We Must Regulate A.I. Here's How,"
implies. You wrote that "A.I. tools are being trained on huge troves of data in ways that are
largely unchecked. Because they may be fed information riddled with errors and bias, these
technologies risk automating discrimination--unfairly locking out people from jobs, housing, or
key services."3

For the FTC to undertake new regulation or an investigation, more than fearmongering
and fanciful speculation are required by law. The FTC Act requires that the Commission have a
"reason to believe" that a party possesses evidence of an unfair or deceptive act or practice in

1 Briefing by FTC Staff to Committee Staff (June 2, 2023).
2 See Sam Altman (@sama), X (July 13, 2023, 5:24 PM), https://twitter.com/sama/status/1679602638562918405; Cat Zakrzewski, FTC investigates OpenAI over data leak and ChatGPT's inaccuracy, WASH. POST (July 13, 2023), https://www.washingtonpost.com/technology/2023/07/13/ftc-openai-chatgpt-sam-altman-lina-khan/.
3 Lina Khan, We Must Regulate A.I. Here's How, N.Y. TIMES (May 3, 2023),
https://www.nytimes.com/2023/05/03/opinion/ai-lina-khan-ftc-technology.html.

2

order to issue a CID. 4 Your op-ed argues for going after "not just the fly-by-night scammers" but
also "the upstream firms that are enabling them" by producing problematic AI "tools."5 This
approach is a stark departure from past FTC practice, as the Commission has traditionally
focused on the harm caused by a product's use--not its design--in its enforcement actions.
Furthermore, such regulation would represent an astonishing expansion of power over otherwisebenign products. It would be akin to the FTC regulating a cell phone's design in order to enforce the do-not-call registry.

Your comments were reinforced by FTC staff during a subsequent briefing to the Committee about AI on June 2, 2023.6 During the briefing, FTC staff made clear that the agency is looking for ways to determine if data sets used to train AI models are biased, discriminatory, or contain "misinformation," suggesting the FTC was considering an expansive regulatory approach to AI to crack down on non-commercial speech. Your staff's response to concerns that the FTC would, in assessing bias or misinformation, be operating outside its statutory authority and acting as "speech police" for broad swaths of data were vague and unsatisfactory.

While the FTC undoubtedly has the statutory authority to initiate enforcement actions
against companies engaged in "unfair or deceptive acts or practices," the FTC may not launch a
preemptive regulatory approach against code underlying AI systems in order to prevent "bias" or
preclude the use of undefined "discriminatory" datasets. Such an extralegal approach would
inevitably involve the policing of constitutionally protected speech, including the internet or
user-derived data used to train AI models. This is well beyond FTC's statutory mandate. The
FTC has no authority or business attempting to regulate constitutionally protected speech.

Given this context, the CID that the FTC sent to OpenAI is particularly troubling, as is
the fact that the CID was leaked. As Sam Altman, CEO of OpenAI, noted, such a leak "does not
help build trust" between the company and government regulators.7 Moreover, the questions and
document requests within the CID suggest that the FTC is now implementing many of the
alarming legal theories that senior agency leaders told Committee staff that they were
contemplating. The CID seeks information on the training data for OpenAI's Large Language
Model, such as the content categories and languages incorporated.8 The CID also asks about
instances where ChatGPT has led to the "safety challenges" identified in OpenAI's GPT-4
System Card, which include "harms of representation" and "disinformation."9 To the extent it is
even constitutional for Congress to prohibit such speech-based harms, Congress has not done so
here nor authorized FTC to pursue these issues. Finally, the CID directs OpenAI to snitch on
users of ChatGPT who engineered prompts to circumvent ChatGPT filters and rules, a new form
of surveillance with the disturbing potential to chill free speech.

4 15 U.S.C. § 57b--1(c)(1).
5 Khan, supra note 3.
6 Briefing, supra note 1.
7 Altman, supra note 2.
8 Civil Investigative Demand, FTC File No. 232-3044 at 5.
9 Id. at 12; OpenAI, GPT-4 System Card, 4 (Mar. 23, 2023), https://cdn.openai.com/papers/gpt-4-system-card.pdf.

3

So that I may better understand the FTC's views on its regulatory and enforcement
authority with respect to AI, please provide written answers and documents responsive to the
following questions no later than September 25, 2023.

1. What factors do you believe should prompt the FTC to shift from reviewing AI outputs to
analyzing data inputs for algorithmic bias, discrimination, or misinformation?

2. Does the FTC plan to evaluate either data used for training Large Language Models or
the sources of such data for bias, discrimination, or misinformation?

3. How does the FTC plan to identify and address bias, discrimination, or misinformation
within diverse training datasets that incorporate content from both commercial and noncommercial speech sources, as well as from user interactions with AI models? Detail the
FTC's specific technical and legal approach.

4. Does the FTC have statutory authority to review prompts submitted by users to
generative AI systems? If so, please describe the statutory basis for that authority and
how FTC intends to or can use those data.

5. How did the FTC's CID to OpenAI leak? In answering, identify the source of the
apparent leak, if known, and detail the steps you are taking to investigate and address the
leak.

6. Detail the approval process for the CID issued to OpenAI, identifying the signatory,
clarifying whether it received a Commission vote, and confirming if it falls within the
scope of an adopted omnibus resolution; if the latter is the case, provide the resolution.

7. Did the FTC possess evidence before issuing the CID warranting a reason to believe that
OpenAI violated Section 5? If so, provide documentation sufficient to show FTC's
reason to believe there was a violation.

8. The CID demands that OpenAI detail measures related to filtering or blocking inputs and
outputs of its Large Language Models.10

a. How do OpenAI's efforts to control model inputs or outputs relate to an alleged
Section 5 violation?

b. Does the FTC expect OpenAI to implement input/output filtering to comply with
Section 5, and if so, what would such measures entail?

10 Civil Investigative Demand, supra note 7 at 9.

4

c. In light of the FTC's discussion of "inoculation theory" in a June 2022 report,11 will the agency apply this theory in evaluating OpenAI's Large Language Models, specifically by assessing the use of measures, such as "prebunking" or "debunking," designed to counteract "online misinformation?"

Thank you for your attention to this matter.

Source