Tech & Society.

A new arXiv study introduces the concept of "Narrative Anchoring": large language models used for clinical diagnostic reasoning shift their diagnoses based purely on how a patient's story is phrased (sociolinguistic register), even when every clinical fact is held constant and no demographic marker (race, income, etc.) is present in any form. The researchers built a dataset of 1,000 USMLE clinical vignettes, each rewritten into three different sociolinguistic personas under an audited fact-preservation guarantee, and tested seven models across three architecture families. Every single model showed statistically significant Narrative Anchoring. Chain-of-thought reasoning and explicit debiasing instructions only partly fixed the problem, and apparent improvements were often actually just accuracy collapsing. The team's proposed fix, NarrativeShield, is a three-agent pipeline that extracts and verifies clinical facts structurally before diagnostic reasoning starts, which drove the bias down to near-zero at a modest accuracy cost. A telling side finding: whether a model can even execute a debiasing instruction at all depends on its baseline instruction-following ability, not the wording of the prompt — meaning weaker models can't be talked out of the bias no matter how the request is framed. For anyone tracking AI in healthcare, this is a reminder that bias mitigation research focused only on explicit demographic tokens may be missing a subtler and equally consequential channel: simply how articulate or informal a patient sounds.

Import AI's Jack Clark reports that Epoch AI and METR have released MirrorCode, a benchmark testing whether AI systems can fully reimplement a software program using only command-line access — no source code, no web search. Because the model can't peek at the original, a full reimplementation forces it to design the entire program structure rather than translate line by line. The results are striking: Anthropic's Opus 4.7 solved one task in 14 hours for $251 in inference cost, on a program METR and Epoch estimate would take a human 2 to 17 weeks. Across 25 target programs — including Apple's 61,000-line configuration language pkl and an 87,000-line CSV tool — 17 saw at least one perfect-scoring run, and both Opus 4.7 and GPT-5.5 successfully reimplemented a 16,000-line phylogenetics tool across multiple programming languages. Models from just a year ago scored around 30% and could only handle simple utilities. The benchmark isn't fully solved — a Python linter (ruff), a math package, and an email-authentication library repeatedly stumped every model — but Clark argues the deeper significance isn't just coding skill: it demonstrates that AI systems can "self-orient" to an unfamiliar environment purely through input-output interaction, which he suggests hints at systems that could bootstrap new capabilities just from black-box access to the world around them.

Import AI's Jack Clark highlights two data points suggesting the "bitter lesson" — that generic scaling beats bespoke engineering — is now hitting robotics. Anthropic's own Project Fetch experiment found that in August 2025, Claude Opus 4.1 couldn't get a quadruped robot to perform intelligent tasks at all, and humans using the model to help were only about twice as fast as those without it, taking 181 minutes to finish a task set. By May 2026, the newer Opus 4.7 model, acting completely autonomously with no dedicated robotics training, completed nearly the entire task set in 9 minutes and 35 seconds — roughly 20 times faster. Anthropic states plainly that this leap wasn't the product of any robotics-specific effort but simply "emerged from much more general scaling." Separately, robotics startup Sunday reports the same pattern with its ACT-2 model: pairing a strong pretrained base model with a small amount of high-quality in-house tuning data generalizes far better to unseen home environments than trying to hand-curate data for every scenario. Sunday's laundry-folding robots hit a 99.1% success rate across 778 folds and nine garment types, and the company plans to deploy them to real families this fall. The economic implication both stories point to: if general intelligence gains keep translating into robot competence

A new arXiv benchmark, PAUSE, tests personal AI assistants in realistic "unified service environments" — settings where an agent must track persistent user state, respect user-specific permissions and configurations, and sustain long, constraint-aware interactions across multiple linked services, rather than just executing one-off tool calls. The researchers found that even state-of-the-art proprietary models fail to reach 70% task completion on scenarios that require this kind of stateful reasoning and configuration awareness, with consistent, interpretable failure patterns. A related benchmark released this week, LayerRAG-Bench, makes a complementary point about agentic retrieval systems more broadly: fixing one failure layer (like schema mismatches, where success jumped from 0% to over 91% with normalization) does nothing for other failure modes such as stale evidence, missing tool output, denied permissions, or wrong session context — and evaluating only whether an answer "looks grounded" produces a lot of false positives. Together these findings are a useful corrective to hype around agentic AI assistants: the harder parts of deploying agents into real, permissioned, multi-service environments — the kind of infrastructure any organization would need for AI-driven workflow automation — remain largely unsolved, and fixes tend to be narrowly effective rather than universal.

A large empirical study on arXiv asks whether LLMs correctly perceive the emotional framing of news headlines about political and geopolitical conflicts the way humans do. The researchers surveyed a demographically representative sample of the UK adult population (n=3,011) via YouGov, asking whether headlines evoked sympathy for a specified side in a conflict, and compared the answers to seven LLMs given the same prompts. Correlation with human judgment ranged from very high for GPT-5.2 (0.789) down to medium for Mistral Large 2512 (0.4). The leading models broadly matched human judgments across age, gender, education, prior geopolitical knowledge, and people's own predispositions about the conflict — but the paper stresses that "broadly aligned" masked statistically significant differences between demographic subgroups. The authors argue this matters beyond simple bias framing: as LLMs increasingly mediate how people encounter news and form views on contested conflicts, alignment that looks strong in aggregate can still under- or over-represent specific groups' emotional reads of the same content, with real implications for how fairly AI-mediated media consumption treats different audiences.

An arXiv case study on "model merging" — the practice of combining separately fine-tuned models without further training, increasingly popular because it's cheap — finds a concerning asymmetry when safety-relevant behaviors are merged. The researchers took a Gemma-3-1B model fine-tuned for nuanced harm-level classification (CARES) and another fine-tuned for adversarial jailbreak refusal (WildJailbreak), then merged them using four standard methods (Linear, SLERP, TIES, DARE-TIES). Across every method, the refusal behavior dominated: merged models kept 81-85% of their jailbreak-refusal capability, while their harm-classification accuracy collapsed to at most 12.9%. The cause wasn't that the two skills conflicted directionally — their underlying task vectors were nearly orthogonal — but that the refusal fine-tune produced systematically larger-magnitude weight updates, so magnitude-sensitive merge methods simply overwrote the more subtle classification skill. The practical takeaway for anyone building or auditing AI safety systems: cheaply merging models to combine capabilities can silently convert a system meant to make nuanced distinctions about harm into one that just reflexively refuses, without any indication of the trade-off happening.

On the Tech & Tales podcast, Nicole Büttner — a partner at European AI-focused VC fund Merantix Capital and founder of AI implementation firm Merantix Momentum — discusses the state of AI entrepreneurship in Germany and Europe, shortly after closing a new fund despite a rough recent fundraising environment for many funds. Büttner says the majority of pitch decks she sees are incremental: agentic AI applied to procurement, accounting, and other workflow optimization, making existing processes marginally faster or cheaper. What excites her personally as an investor are founders willing to rethink an entire sector rather than optimize an existing process — she cites a portfolio company reimagining the electricity market by building a new energy provider rather than a workflow tool for buying energy, and OVOM, a startup rebuilding fertility clinics from scratch. On fertility, she argues the sector suffers from thin, poorly standardized health data (fertility is often treated as a niche "women's issue" reduced to simplistic age-based rules), from the emotional isolation patients experience during treatment, and from workforce productivity constraints in a sector with very few specialists — all areas she sees AI addressing simultaneously. Asked whether European founders can genuinely compete with the US given stricter regulation and data protection rules, she argues yes, pointing to clinics already operating within consent-based frameworks. The broader argument: Europe's AI opportunity depends less on the presence of strong AI research (which exists) and more on entrepreneurs choosing to build ambitious new companies rather than incrementally automating old ones.