Latest / AI Ethics with Fexingo: Bias, Safety, and Responsible Artificial Intelligence / AI Transparency Reports Promise More Than They Deliver
Transcript
- Lucas: So you've probably seen the headlines — OpenAI releases a new 'system card,' Google publishes a 'model card' for Gemini, Meta puts out a 'responsible AI report.' These glossy PDFs are supposed to tell us how the models work, what biases they have, what safety testing was done. But here's the uncomfortable question: are these transparency reports actually making AI more accountable, or are they just sophisticated PR? Luna: I've read a few of them. They're dense. Lots of technical jargon, performance benchmarks... but I always feel like I'm missing the real story. What's actually being hidden? Lucas: That's exactly the tension we're going to unpack today. Because transparency reports — especially the ones called 'model cards' — were originally proposed by researchers at MIT and other institutions as a way to standardize disclosure, like a nutrition label for AI. The idea was you'd know the ingredients, the risks, the testing procedures. Luna: Right. And in theory, that's great. If every company published a model card, regulators, researchers, and users could compare apples to apples. Lucas: But what's happened in practice is more like a restaurant publishing a menu that lists 'beef' without telling you the cow was fed antibiotics or raised on a deforested patch of the Amazon. Let me give you a concrete example. Anthropic, the company behind Claude, publishes model cards. Their most recent one for Claude 3.5 Sonnet is about 40 pages. It details the model's architecture, training compute, benchmark scores on math and coding, even some red-teaming results. Luna: Sounds thorough. So what's missing? Lucas: Two critical things. First, the training data. The model card says the data is 'a mixture of publicly available and licensed data' — that's it. No specifics about what percentage is web-scraped versus licensed, no breakdown of domains, no information about how they filtered for bias or toxicity. This is the single biggest source of risk in any large language model — the data — and it's described in one vague sentence. Luna: They probably argue it's proprietary. If they revealed their data sources, competitors could copy their advantage. Lucas: Exactly the argument they make. But here's the thing — the MIT researchers who pioneered model cards explicitly called for 'dataset characteristics' to be included. They wanted to know the size, the sources, the collection methodology, the language distribution. Instead, what we get is a marketing gloss. And the second missing piece: details on red-teaming methodology. Anthropic says they did 'red-teaming evaluations.' But they don't tell you who the red team was, what scenarios they tested, or what failures they found. Luna: So it's like a bank saying 'we stress-tested our portfolio' but not telling you what interest rate shock they assumed or whether they passed. Lucas: Exactly. And this isn't unique to Anthropic. OpenAI's system card for GPT-4 Turbo runs over 60 pages. It lists a huge range of evaluations — bias, factuality, refusal rates, dangerous capabilities. But again, the training data section is opaque: 'a large-scale corpus of text and images from the internet.' And the red-teaming section describes the general process but gives no raw numbers on how many test cases, what categories of attack, or what the worst-case outcomes were. Luna: I remember one eyebrow-raising detail from that report. They disclosed that the model sometimes 'mirrors user bias' — but they didn't quantify how often, or in what contexts. That feels like a missing piece for someone trying to decide whether to deploy this model in a medical setting, for example. Lucas: Right. And that's the core problem: transparency reports are currently designed to protect the company from liability, not to empower the end user. A doctor who wants to use GPT-4 to summarize patient records needs to know: does this model have higher error rates for non-white patients? For non-English speakers? For certain medical conditions? The system card won't tell you. Luna: So what would a genuinely useful transparency report look like? Can you give me a model — a hypothetical or a real one — that does it better? Lucas: There is one interesting example from a non-profit: the 'Model Card' for the BLOOM model, which was created by the BigScience project. BLOOM is a large language model built by an international research collaboration. Their model card is roughly the same length as Anthropic's, but it includes a data sheet that lists every dataset used, including size, language, domain, and even potential biases in each source. It also includes a section on 'intended use' and 'out-of-scope use' with specific examples. And they published the results of a bias evaluation broken down by demographic group — not just aggregate scores. Luna: So the difference is that BLOOM's transparency report actually empowers a downstream developer to make an informed decision. Anthropic's report gives you warm feelings but not actionable data. Lucas: That's exactly the distinction. And it's not just about model cards. Google publishes 'AI Principles' and transparency reports on their products. In 2025, they released a report on 'responsible AI' that included a section on their speech recognition model. They disclosed that the word error rate was 4.9 percent overall — but when they broke it down by dialect, it was 3.8 percent for General American English and 11.2 percent for African American Vernacular English. That's a threefold increase! That's the kind of specific disclosure that lets researchers and advocates push for improvement. Luna: And that kind of data is actionable. You can demand better training data for underperforming dialects. You can measure progress over time. Lucas: Exactly. But the problem is — those numbers only came out because of sustained pressure. The default for most companies is to publish aggregate metrics that look good and bury the disparities. So the question becomes: how do we shift the norm from 'transparency as PR' to 'transparency as accountability'? Luna: I think regulation is the obvious lever. The EU AI Act, which is starting to take effect now in 2026, requires certain high-risk AI systems to disclose information about their training data and performance. But I've heard critics say the requirements are too vague to force real disclosure. Lucas: You're right. The EU AI Act does require documentation for high-risk systems — things like medical devices, resume screening tools, credit scoring. But the language is open-ended: 'a description of the training data sets used, including their origin, scope, and main characteristics.' It doesn't mandate the level of detail that researchers need. And it has an exception for trade secrets, which is a huge loophole. A company could argue that revealing its training data sources would reveal proprietary curation methods. Luna: So we're back to the same standoff: transparency versus intellectual property. Is there any way to reconcile them? Lucas: I think there is, but it requires a different approach. Instead of demanding full disclosure of training data, regulators could demand disclosure of evaluation results at a fine-grained level, broken down by demographic groups, languages, and task categories. That way, companies can keep their data proprietary, but we can still see how the model performs across different slices of the population. Luna: So you're saying — keep the recipe secret, but publish the nutritional content per serving. Lucas: Exactly. And that nutritional content should be verified by an independent auditor, not just self-reported. A few startups are already doing this — there's a company called Credo AI that audits models and issues third-party model cards. But they're not yet a standard requirement. Luna: It feels like we're in a transition period. The first generation of transparency reports was essentially marketing. The second generation, maybe the one we're entering now, could be more substantive — if regulators and customers demand it. Lucas: I think that's right. And there's a market incentive too. When Google disclosed that dialect gap in speech recognition, they took a PR hit — but they also committed to fixing it, and they've since narrowed the gap. Companies that embrace real transparency could build trust. And in a world where AI is being deployed in high-stakes decisions — hiring, healthcare, criminal justice — trust is a competitive advantage. Luna: But do you think most companies see it that way? Or are they still in the 'hide the problems' phase? Lucas: I think the smart ones see which way the wind is blowing. The EU AI Act is just the beginning. California is considering its own AI transparency law. The White House has issued executive orders on AI safety that include reporting requirements. Over the next three to five years, transparency reports are going to shift from voluntary glossy PDFs to regulated, standardized documents. The companies that start now by being genuinely transparent — like what BigScience did with BLOOM — will have a head start. Luna: So the question for our listeners, especially if you're a developer or a product manager using AI, is: are you reading these reports critically? Are you asking for the granular data before you deploy a model? Lucas: That's the challenge. Because as of today, a model card often tells you more about the company's marketing department than about the model's real risks. The next time you see a transparency report, look for three things: first, a data sheet that actually describes the training data sources and their known biases. Second, evaluation results broken down by subgroup — not just the overall accuracy. And third, a clear description of red-teaming methodology and findings. If those are missing, you're not getting transparency — you're getting PR. Luna: And if you're building a product on top of one of these models, that's not just an ethical risk — it's a business risk. You could be building on a biased foundation without knowing it. Lucas: Exactly. And when the regulator or the lawsuit comes calling, 'we trusted the model card' won't be a defense. So for now, the burden is on all of us to read critically. But I'm hopeful that in a few years, a real transparency report will be as standard and as useful as a nutritional label on your cereal box. Luna: A cereal box that actually tells you how much sugar is in each serving. Not just 'made with real fruit.' Lucas: Precisely. So next week, we're going to dig into another angle of AI accountability — the growing ecosystem of AI auditors and what they actually find when they poke under the hood. Thanks for listening.