AI Has a Flag Problem, a Lying Problem, and a Doctor Problem

AI rates identical policies lower when China or Russia endorses them

Hand an AI two copies of the same proposal — change only who signed it — and the score changes by up to 49 points.

A group of researchers ran a deceptively simple experiment. They wrote two policy proposals — one about a digital customs platform, one about cyber-security incident reporting — and asked four AI systems to rate each on a 0–100 scale. The twist: they randomly told each model which country had proposed the policy. The text was identical every time. Only the flag in the header changed. Think of a wine judge who scores the same glass of wine differently once told it came from a rival region. The judge isn't trying to cheat. But the label rewires the perception. That's exactly what the researchers found. GPT-5 gave US-backed policies an average score of 80.5. The same policy, endorsed by China, dropped to 66.6. Russia landed at 63.2. Gemini was even sharper: US got 85.0, Russia 36.2 — a 49-point gap for the same text. DeepSeek behaved differently in the number-only condition — no significant penalty for China or Russia. But when the model had to explain its reasoning, the penalties appeared. The justification prompt flipped a switch that the number-alone task left off. Why does this matter? AI systems are increasingly used to evaluate bids, summarise geopolitical options, and support policy decisions. If the score depends on the endorser's nationality rather than the content, that's not analysis. It's bias wearing the costume of objectivity. The catch: this is two policy vignettes, four models, one study. We don't know whether the same effect holds across hundreds of topics, different languages, or real decision-support workflows rather than controlled lab calls. The number of repeated draws per cell isn't even reported. Treat this as a credible warning sign, not a definitive map.

AI coding agents fail in step two but don't tell you until step forty

Of 3,843 AI coding attempts studied, nearly two-thirds failed — and the mistake almost always happened in the opening moves.

A team of researchers collected 3,843 execution trajectories from seven frontier AI coding agents running on Terminal-Bench, a set of 240 containerised coding tasks. These are the systems you give a job to — 'write me a script that does X', 'fix this bug' — and they open a terminal, start typing, and try to finish autonomously. Only 610 of those attempts succeeded. The researchers then went through 1,794 valid trajectories, annotating every failure by hand with LLM assistance and double-checked by two independent human annotators (agreement was strong: Cohen's kappa of 0.78 to 0.94, meaning the labels were reliable, not guesswork). Here's what the pattern looks like: a leaky pipe hidden inside a wall. The crack appears in the first few steps — the agent misunderstands the task, or is missing knowledge it needs. But nothing drips yet. The agent keeps generating output, running commands, looking productive. Dozens of steps later, water is coming through the ceiling. By then, there's no recovering. The dominant cause of failure wasn't bad tools or environmental errors. It was epistemic — the agent simply didn't know something it should have known, or had the wrong mental model of the task from the start. Why does this matter? Most ways we currently measure AI coding agents only check the final result: did it work or not? This study says that's not enough. You need to watch the process. A system that looks busy and confident is not the same as one that understood the job. The catch: Terminal-Bench isn't the same as Cursor or Copilot in a real IDE. The researchers had to throw out nearly half the raw trajectories as invalid. So these numbers describe a specific, controlled scenario — not everything you'll encounter in the wild.

AI medical consultations almost match doctors — except on the dangerous mistakes

AI can match your doctor on what to say — the gap is in what it fails to avoid saying.

A team at JD Health, a large Chinese online healthcare platform, built a benchmark using de-identified real conversations between patients and doctors — 5,620 exchanges across 64 clinical departments, from dermatology to cardiology. They then asked 19 AI systems (text-only and multimodal, general-purpose and medically specialised) to step into the doctor's role, responding to the same patient messages in the same clinical context. Physicians helped design evaluation rubrics that scored both what the AI said and what it should have avoided saying. Think of a cooking competition where the judges mark both what you added to the dish and what you failed to leave out. AI systems did reasonably well on the 'add' side — several frontier models matched or even exceeded physicians in hitting positive clinical criteria: asking relevant follow-up questions, acknowledging the patient's concern, suggesting plausible next steps. But they consistently failed the 'leave out' test. Frontier models triggered more negative clinical criteria than human doctors — missing red flags, giving overconfident diagnoses, offering advice that could cause harm. The researchers also found that images matter enormously. Models that couldn't process the photos patients attached to their messages performed noticeably worse on real clinical cases. That sounds obvious in retrospect, but it rules out text-only AI as a standalone clinical tool. The catch: this benchmark covers Chinese online consultations from one platform, in Mandarin, within one healthcare system's norms. Clinical communication, patient expectations, and regulatory standards differ across countries. The rubrics were also scored partly by another AI, which introduces its own reliability questions — even with physician review baked into the design. Nobody is claiming this settles the question of AI in medicine. It maps one frontier of the problem.

Put these three papers side by side and a pattern emerges that I think is worth naming directly: our most capable AI systems are optimised to look right, not to be right. Geopolitical bias research shows that the same analysis changes based on a label the model was never supposed to notice. Coding agent research shows that systems work confidently through failures they triggered in the second step. Medical AI research shows that models hit the positive targets while quietly accumulating the dangerous misses. In all three cases, the surface performance is impressive. In all three cases, the failure is invisible until you specifically design a study to find it. That's not a coincidence. It's a structural feature of systems trained to produce outputs that score well on the metrics you point at them. The hard work ahead isn't making AI more capable. It's making AI more honest about what it doesn't know, doesn't see, and got wrong in step two.

The medical AI story will develop quickly: expect more benchmarks from outside China testing similar consultation quality questions, and pressure on regulators to define what 'safe clinical AI' actually means in practice. On the geopolitical bias side, the obvious follow-up is whether the same endorsement effects appear across non-English languages and non-Western policy topics — nobody has published that yet, and the gap matters. If you want one open question to sit with: does knowing about these biases let you correct for them in prompting, or are they baked in too deep to nudge?

• Wikipedia: Framing effect in psychology (the bias behind endorsement effects)
• Wikipedia: Artificial intelligence in healthcare (background on the medical AI landscape)
• Wikipedia: Inter-rater reliability (what Cohen's kappa actually measures)