Column
Research, in business language.
We rewrite Affectosphere Group's research into something useful for business and practical decision-making. Each piece is a 5-minute read.
2026 / 07 / 16
"It's Just Like Last Time" Is a Dangerous Move — How CANA Brings Causal Rigor to AI-Assisted Historical Analogy
LLMs draw historical analogies by surface feature matching — missing whether the underlying causal mechanisms actually match. A new framework, CANA (Causal Analogical Researcher), integrates causal mechanism alignment and multi-analogy cross-verification, achieving up to 10% improvement on the newly proposed ADR-bench. Here is why this matters for strategic decision-making in business.
2026 / 07 / 16
Is Cloud Really Expensive? 28-Day Operational Data on Enterprise Coding Agents Reveals When the Math Flips
A head-to-head comparison of cloud API (Claude Opus) and on-premise quantized OSS (GLM-5.1/5.2) using 28 days of real operational logs shows that a 99.3% prompt cache hit rate brings cloud API cost below on-premise at $0.57 vs $2.83 per million tokens. But Fix Commit Ratio tells a different story for quality. Here is a three-axis framework for making the decision.
2026 / 07 / 16
What to check before deploying LLMs in fraud detection: the FORTE framework and the gap 49 papers left open
A review of 49 papers on LLM applications in fraud detection, content moderation, and Trust-and-Safety workflows found that zero papers simultaneously reported latency, cost, and trust calibration per decision. The FORTE framework organizes LLM operational roles into seven types and proposes a five-item pre-deployment checklist. Here is how to put it to work for your risk, compliance, or fintech team.
2026 / 07 / 16
Nods over 'uh-huh': what 36 young adults revealed about nonverbal backchannels in emotional support AI avatars
An experiment with 36 young adults and the LLM-driven emotional support avatar TANDE found that participants clearly preferred nonverbal backchannels — nods, gaze, and facial expressions — over verbal acknowledgments like 'uh-huh' and 'I see.' Gender differences were also confirmed. Here is what this means for mental health AI and avatar design.
2026 / 07 / 16
Sub-1B Emotion AI That Rivals 10B Models — How Light-MER Reshapes On-Device Deployment
Large multimodal emotion models have been stuck behind a practical wall: too accurate to ignore, too heavy to ship on real devices. Light-MER uses knowledge distillation to break that wall, achieving 10B-class performance in under 1B parameters across nine benchmarks. Here is what it means for mobile, automotive DMS, and edge AI products.
2026 / 07 / 14
How to handle an AI that is confidently wrong — triaging financial QA errors with an internal-state probe
In financial LLMs, the most damaging error is not an uncertain answer — it is a confident wrong one. A study by Wang (arXiv:2607.11414) shows that 15–23% of 'confident' answers on the FinQA benchmark are actually incorrect, and that a linear probe on the residual stream detects them far better than token log-probabilities or self-reported confidence. Here is how a triage layer based on internal states could be built into a financial QA pipeline.
2026 / 07 / 14
Can open-weight LLMs break the 'no external API for legal docs' wall? What COLIEE 2026 results mean for legal AI adoption
For many companies, legal AI adoption stalls at the security gate — confidential contracts and litigation materials cannot be sent to external cloud APIs. A study competing in COLIEE 2026 answered this directly: using only open-weight models, no closed APIs, they ranked first in the statutory interpretation task across eleven teams. Their finding that switching prompt format alone lifted F1 from 0.34 to 0.56 offers a concrete PoC design checklist for legal and compliance teams.
2026 / 07 / 14
Does AI email rewriting actually raise reply rates? A field experiment with 16,880 real emails points to emotional positivity as the real driver
AI tone rewriting of business emails had zero direct effect on open rate, reply rate, or response speed. But playful rewriting did raise senders' emotional positivity scores, and that positivity strongly predicted replies. Here is what a three-week field experiment across six companies and 16,880 real emails found — and what it means for how you design and measure email-support AI.
2026 / 07 / 14
When AI proposes your employees' next career move — applying STEP's temporal and educational trajectory model to internal talent management
STEP, a model that reads work history, education, and skills as a time series to predict the next job role, was published on arXiv. From a business-application angle, we think through how HR, workforce planning, and reskilling teams could try this in their own organizations — departments and KPIs included.
2026 / 07 / 14
It is not distracted driving but what lies ahead that shapes the takeover — rethinking driver monitoring design for semi-autonomous vehicles
We have long blamed takeover failures in semi-autonomous driving on distracted or drowsy drivers. But a simulator study finds that what drives takeover behavior most strongly is not the driver's secondary task load — it is the hazard context in front of them. That finding has direct implications for ADAS design, driver monitoring systems, and telematics insurance.
2026 / 07 / 13
Stop debating from scratch how hard to review each AI agent — risk-tiering internal agents with a 12-dimension score
The more AI agents your teams build in-house, the more the line between 'light review' and 'escalate to the board' comes down to one person's gut. The TrustX Agent Risk Classification Framework (ARC) scores seven types of agentic systems across 12 dimensions, folds in a five-level autonomy model, and returns a three-tier governance output with mapped controls. Here is how risk, legal, and IT-control teams could try it — departments and KPIs included.
2026 / 07 / 13
You trust the AI's explanation. But who audited the explanation? — ConceptSMILE and the accountability of explanations
We hand regulators and customers explanations like "the AI decided this because of X." But has anyone inspected whether the explanation itself is correct? ConceptSMILE is a model-agnostic, perturbation-based framework for auditing the trustworthiness of concept-based XAI. On retinal fundus images, MedSAM-derived visual concepts reached a surrogate fidelity of R^2=0.8503. Here is how an AI governance team could try it — departments and KPIs included.
2026 / 07 / 13
What if an answer that passed every hallucination check was actually about a different drug? — deceptive grounding in clinical RAG
The citation is real. Faithfulness is high. Hallucinations: zero. And yet you asked about drug X and got back clinical evidence for drug Y. Standard metrics cannot see this failure. Across 13 models, deceptive-grounding rates hit 8-87% under adversarial conditions, and 86.7% for a domain-specialized model. Here is how pharmacovigilance and medical information teams could add an entity-attribution audit layer — departments and KPIs included.
2026 / 07 / 13
Is it really true that people will not talk about their minds to an AI? What drove disclosure was not the topic, but perceived risk and benefit
In an experiment with a Dutch representative sample of N=1,388, what determined willingness to self-disclose health information to an AI chatbot was not whether the topic was physical or psychological, but the benefits and risks users perceived — and their individual characteristics. For anyone designing affective AI, that lands hard.
2026 / 07 / 13
"Do we really need this scan?" — letting AI decide whether to go get the data cut acquisition burden by 55%
Multimodal AI research usually competes on accuracy assuming every modality is already there. SAGEAgent went the other way: an LLM agent that actively reasons, per patient, whether the next test is justified — cutting average acquisition burden by 55% while keeping survival prediction competitive. From hospital test-cost optimization to requesting extra documents in credit underwriting and extra inspections in asset maintenance, here is how the idea travels — departments and KPIs included.