This has been a week of exceptional BS in AI-land from "superintelligence" to "dangerous AI" (dangerous disembodied tech mind you! not dangerous companies with deeper pockets and more market power and entitlement than anyone's ever seen!!) to yet another round of wind-bloated predictions about AI's economic impact... In the middle of all of this, we thankfully do have some terrific foundational scholarship to draw on. For example, "AI auditing: The Broken Bus on the Road to AI Accountability" by Abeba Birhane, Ryan Steed, Victor Ojewale, Briana Vecchione, PhD, and Deborah Raji is excellent context to better understand what is and is *not* happening when OpenAI or Anthropic asks an external organization to review their security failures. Open-access version available here: https://lnkd.in/ggbzADFQ
Peter Hill
1mo
This article is inaccurate. To say the practical nature of the "AI audit" ecosystem is muddled and imprecise is made-up or ill informed.
Amitava Sarkar
1mo
I agree that most of AI "news" is deceptive - and fluff - and attempts of hedge. However, anyone claiming to be able to "audit" AI raises skeptical flags for me. Audit implies an agreed to boundary/scope. It implies measurements and other aspects. AI - Intelligence remain undefined. Benchmarking - no standards - people should go explore how differently each AI company benchmarks - yet cannot tell me the cost of a prompt (P.S. It changes the longer the session). Not a single release note - when any of the slaves processes around the LLM - or even censorship - changes . They are changing the behavior of the AI engine more often than they will even admit. Opus 4.8 on day 1 is different from Day 100 to Year 1. Have how many release notices? What will be audited when? and be how reliable/meaningful? However this is somewhat of a topic that interests me- and will let AI analyze the messages, in the pdf. Thanks for post
The accountability question deserves far more attention than another round of predictions about AI capabilities.In enterprise AI, were seeing a similar gap between demonstrating what a system can do and explaining why it made a particular decision, what evidence it used, and who is accountable when it gets things wrong. At Colrows, we believe that business context and traceability need to be built into AI systems from the start. Auditing after the fact can only tell us so much.
Razvan Savin
1mo
Independence is the part most teams fake. The paper only counts something as an audit if the checker is adequately separate from the builder, or if the check is a distinct process from engineering. That second clause matters. A review can be useful and still not be accountability. If the target defines the scope, controls the relationship, and narrates the result, the audit has no teeth. Different domains keep hitting the same problem: evidence is not enough unless the process gives it consequence.
Mohamed Elareef
1mo
Most published AI audits still sample model outputs, not the compute contracts and procurement clauses that concentrate the power.
Marc Richardson
1mo
Thanks for sharing this article. I wonder if the use of the term "AI Accountability" is still giving too much focus to the "AI" rather than the companies building it. To a bystander, this sounds like, "how do we make these systems more accountable" rather than "how do we make these people more accountable"
Mikhail Shakh
1mo
AI safety gets weaker when we treat technical risk and institutional accountability as competing narratives. A model can have real dangerous capabilities and company deploying it can also have incentives and market power that deserve scrutiny. External audits matter, but only if auditors have meaningful access that turns findings into consequences.
Right there with you on the specificity point. It's always nice to have a BS-detector or at least a pair of BS-vision glasses. I've been building something in that spirit, it's a tool that sheds light on the shadow benchmakrs cast, more specifically, it catches where a model's aggregate accuracy holds up while performance on individual tasks subsets quietly fail or aren't as flattering.
Dario Crespo
1mo
Really interesting. This connects to something we found running an adversarial experiment between models: the outcome of an evaluation depends as much on the protocol as on the model being evaluated. We ran a cross-run where each model designed the other's tests and then evaluated its responses. The most revealing part wasn't the scores, but one specific case: a model ended up evaluating itself without realizing it, and for most of the test assumed it was evaluating a different model , only discovering at the end that it was its own family. The conclusion I keep coming back to: it's not enough to ask , how well did the model pass the audit? ; you also have to audit the mechanism by which you arrived at that conclusion including whether the evaluator actually knows what it's evaluating. Model → protocol → evaluator → score. Each layer can introduce bias or blind spots. I think that's a big part of the accountability problem this paper is getting at.
Prompted also by Andrew Strait's great recent post on the next upcoming "audit" :)