In a wake-up call for high-tech companies developing artificial intelligence products, the Federal Trade Commission (FTC) recently cracked down on a large AI software company that failed to back up AI-related claims as evidence of their actual success. The final order of August 28th is a reminder that long-standing advertising principles apply equally to both traditional businesses and their marketing AI products and services. In this case, what are the five practical lessons that high-tech companies can take from government actions?
The software company has made bold claims about AI products, and the government has notified it
Workado, a software company based in Palo Alto, has developed an AI detector tool that it claims to be extremely accurate. In fact, the marketing material claimed that the tool could identify AI-generated text with 98% accuracy. This argument is particularly pronounced when educators, publishers, and companies are actively seeking reliable ways to distinguish between texts created by generative AI and human-obsessed content.
However, when the FTC investigated Workado's claims more closely, an important issue emerged.
- Training data: Workado promoted the tool as being able to analyze a wide range of content, but the underlying model was actually primarily trained in academic writing such as essays and academic papers, rather than a broader mix of blogs, marketing copies and other online sources represented by the company.
- performance: In testing outside of academic contexts, the accuracy of the tool was reduced to approximately 53%. Effectively, it's far less than pure chance. The FTC called it “not better than coin toss.”
- False: The FTC determined that Workado's marketing had effectively exaggerated the functionality of its products and that these representations misunderstood customers about the reliability of the tool and real-world performance.
FTCs are becoming more intense in high-tech companies
As a result, the FTC approved the final consent order on August 28th.
- Stop billing for unsupported precision. Workado must suspend statements about the validity or “accuracy” of the AI Content Director unless those claims are misleadingly supported by “competent and reliable evidence” at the time the statement is made.
- Keep test data and evidence. We need to maintain documentation of how we establish performance claims, including test data and analysis when related to product effectiveness.
- Notify the customer. We must notify you of consent orders and payments and send you an FTC draft notification explaining the issue to you, including ensuring transparency regarding the tool's revised representation.
- I'll report it to the FTC. Workado is required to provide the government with its annual compliance report for four years.
Five practical lessons from AI companies
Given this FTC order, here are some practical takeaways that can guide you to your AI product marketing approach.
1. Test widely, not narrow.
If the product is used in different domains, the tests should reflect that. A model trained in academic writing may look good in essays, but the results can collapse if the customer uses it in social media content or business reports. Don't assume that “working here” is equal to “working anywhere.”
2. Don't let your marketing team run in front of the data science team.
Ambitious claims often come from the desire to stand out in crowded markets. However, marketing languages need to remain connected to harsh evidence. Practical Steps: Set up a sensual review for reviewing copies for accuracy before technical staff is published.
3. Create an “evidence file.”
All performance claims require a paper trail, including training sets, validation results, methodology, error rates, and limitations. If you are challenging a customer, competitor, or regulatory authority, then the file will be your insurance at your fingertips.
4. Openly review the restrictions.
Some founders fear that careful language when describing AI products will dull “awesome factor.” But paradoxically, I'll admit where your tools can fight increase Reliability. Customers are honest and grateful. “Our model works best with structured texts such as contracts and policies, but may not be as accurate in informal writing” is better than the vague promise of universal accuracy.
5. Building compliance in culture.
You don't need an internal regulatory team to get started. Small practices are a long way to set rules that do not include metrics in public material without routine internal audits, clear versions of claims, and without verification.
