The Problem With 'AI-Assisted': Why Disclosure Isn't Enough
Saying you used AI tells people nothing. Showing how you used it tells them everything.
Every major institution now asks the same question: "Did you use AI?"
Universities put it on submission forms. Law firms add it to engagement letters. Publishers embed it in contracts. The ABA issued formal guidance. ICMJE updated their authorship criteria.
And the expected answer is simple: "Yes, I used AI for [specific task]."
That sounds reasonable. It isn't.
Disclosure tells you nothing useful
Imagine two legal briefs. Both lawyers disclose AI use. But:
- Lawyer A asked AI to "write a summary of relevant case law," copied the output, and submitted it. (This is what happened in Mata v. Avianca. The citations were hallucinated.)
- Lawyer B asked AI "What are the strongest counterarguments to my position?", evaluated each one, dismissed three as inapplicable, and incorporated one into her analysis with her own reasoning.
Both used AI. Both would check the same disclosure box. But the nature of their work is fundamentally different. One delegated judgment. The other used AI to sharpen it.
Disclosure doesn't distinguish between these two scenarios. It treats AI use as binary: yes or no. But the question that actually matters is: how did human judgment shape the result?
The regulatory gap
The EU AI Act, effective August 2026, requires transparency and human oversight documentation for AI-assisted professional outputs. GDPR Article 22 gives individuals the right to explanation for automated decisions.
Neither of these is satisfied by checking a box.
They require evidence. What was the AI asked? What did it return? What did the human do with it? Was the output adopted, modified, or rejected? These questions demand a process record, not a declaration.
What "enough" actually looks like
If disclosure is the floor, what's the ceiling? We think it looks like this:
- Source separation. Every piece of text labeled as human, AI, or human-edited AI.
- Decision tracking. A record of what was adopted, what was dismissed, and why.
- Tamper evidence. An audit trail that can't be edited after the fact.
- Verifiability. A report that a third party can audit without taking anyone's word for it.
That's not a wish list. Those are the four properties that make AI-assisted work defensible. Disclosure covers zero of them.
The bottom line
Saying "I used AI" is not accountability. It's the absence of it. Accountability means showing your work: the decisions you made, the suggestions you rejected, and the reasoning that shaped the final output.
The question isn't whether you used AI. It's whether you can prove your thinking mattered.