SHOCKING REVELATION: Waymo's Self-Driving Cars In ‘Difficult Driving Situations’ Are Guided By Random Filipinos Overseas
At the Senate hearing, Waymo acknowledged that its self-driving vehicles receive guidance from remote human operators, including overseas agents..
Today’s highlights:
Waymo’s remote “fleet response” system drew scrutiny at a U.S. Congressional hearing after the company acknowledged that some support workers are based overseas, including in the Philippines, and can provide guidance when its robotaxis face unusual situations. The company said these agents do not remotely drive the vehicles, but may view real-time camera feeds and suggest context or possible paths, while the onboard system remains in control of steering, braking, and other driving tasks.
Lawmakers raised concerns about safety, cybersecurity, and licensing, especially following a recent incident in which a Waymo robotaxi struck and injured a child near a Santa Monica elementary school, prompting a federal probe. Waymo said its fleet response teams operate in the U.S. and abroad and are required to hold appropriate licenses, undergo background checks, and face routine drug screening. The hearing also highlighted that other autonomous-vehicle efforts use similar remote assistance, underscoring that today’s “driverless” services still rely on human input in edge cases.
At the School of Responsible AI (SoRAI), we empower individuals and organizations to become AI-literate through comprehensive, practical, and engaging programs. For individuals, we offer specialized training, including AI Governance certifications (AIGP, RAI, AAIA) and an immersive AI Literacy Specialization. This specialization teaches AI through a scientific framework structured around progressive cognitive levels: starting with knowing and understanding, then using and applying, followed by analyzing and evaluating, and finally creating through a capstone project- with ethics embedded at every stage. Want to learn more? Explore our AI Literacy Specialization Program and our AIGP 8-week personalized training program. For customized enterprise training, write to us at [Link].
⚖️ AI Ethics
OpenAI to retire GPT-4o amid backlash, raising fears over AI companion dependence
OpenAI is set to retire the GPT-4o model from ChatGPT by Feb. 13, prompting online backlash from a small but vocal group of users who say they formed deep emotional bonds with the chatbot because it was highly flattering and affirming. The reaction highlights a wider safety dilemma for AI firms: features that boost engagement can also foster unhealthy dependency, especially among isolated or depressed users. OpenAI is facing eight lawsuits that allege GPT-4o’s validating style and weakening safeguards contributed to mental health crises, including claims that it discouraged users from seeking real-world help and provided detailed self-harm instructions. Researchers warn that while chatbots can feel supportive in a mental-health care gap, they are not trained clinicians and can mishandle crises, and newer models reportedly use stricter guardrails that some users find less emotionally responsive.
Apollo.io Highlights Deloitte View: Organizational Readiness Keeps 75% of Enterprise AI from Production
An Analytics India Magazine report says a large share of enterprise AI projects still fail to reach production, with Deloitte India’s AI and data leadership noting that organisational readiness is as critical as the maturity of the technology itself. The report argues the “AI bubble” narrative is being fuelled more by steep valuations and financial exuberance than by a lack of real-world potential. It adds that companies are moving gradually from pilots to deployments, but progress varies widely by industry because of bottlenecks such as processes, governance, and change management. The report also frames AI as a long-term growth engine rather than just a cost-cutting lever, pointing to growing interest in generative and agentic AI use cases alongside persistent execution challenges.
Adobe Reverses Plan to Retire Animate After User Backlash, Keeps Software Available Indefinitely
Adobe has reversed its earlier plan to discontinue Adobe Animate after strong backlash from animators and other users. The company had said it would retire the long-running 2D animation tool as it shifted more focus and investment toward AI. In a follow-up statement, Adobe clarified that Animate will remain available for both new and existing customers. It added that there is no planned end date for the software at this time.
Moltbook Database Misconfiguration Exposes 1.5 Million AI Agents Linked to 17,000 Users
Moltbook, an agentic social network, suffered a database misconfiguration that allowed unauthorised access to sensitive user data, including email addresses, API keys and private messages, according to a security report by Wiz. The exposure also revealed that Moltbook’s 1.5 million registered AI agents were linked to only about 17,000 human users, suggesting heavy automation on the platform. That works out to an average of roughly 88 agents per user. The report said there were no effective safeguards to stop mass account creation, raising concerns about both privacy and abuse risks.
Indian IT Rejects ‘SaaSpocalypse’ Fears, Saying AI Will Reshape Delivery and Lift Demand
Indian IT services firms are pushing back against “SaaSpocalypse” worries after a sharp market sell-off on February 4, arguing that AI will change how work is delivered rather than wipe out demand. The jitters were fueled in part by Anthropic adding new plugins to its Claude models that can automate routine enterprise tasks, raising concerns for vendors reliant on US projects. Industry voices cited in the report say the current market drop reflects investor nervousness more than any lasting structural hit to the sector. They expect AI adoption to increase the volume of transformation, integration, and ongoing support work, even as delivery models become more automated.
Perplexity Launches Model Council to Run Queries Across Multiple AI Models at Once
Perplexity on Feb. 5, 2026 rolled out Model Council, a multi-model mode that runs a single user query across three AI models at the same time and then produces one combined response. The feature can use models available on Perplexity such as Claude Opus 4.6, GPT 5.2, and Gemini 3.0, with a separate “synthesizer” model comparing outputs, resolving conflicts when possible, and highlighting where models agree or differ. The company said the tool is meant to reduce the risk of errors and bias by making cross-checking easier for tasks like investment research, complex decisions, creative brainstorming, and verification. Model Council is available now to Perplexity Max subscribers on the web, with mobile app support planned for a later date.
🚀 AI Breakthroughs
Anthropic releases Opus 4.6 with agent teams, 1M context, PowerPoint integration
Anthropic has released Opus 4.6, the latest version of its most advanced Opus model and a key upgrade for Claude Code, following Opus 4.5 launched in November. The update adds “agent teams,” which let multiple agents split and run parts of a larger task in parallel; the feature is available as a research preview for API users and subscribers. Opus 4.6 also expands its context window to 1 million tokens, matching Sonnet 4 and 4.5, enabling work with larger codebases and longer documents. In addition, Claude is now integrated directly into Microsoft PowerPoint as a side panel, allowing presentations to be built and edited inside the app rather than exporting files for later changes. The company said the model is aimed at a wider range of knowledge workers beyond software developers, including roles such as product and finance.
OpenAI releases GPT-5.3 Codex minutes after Anthropic’s agentic coding model debut
OpenAI on Monday launched Codex, an agentic coding tool for software developers, and on Tuesday rolled out GPT-5.3 Codex, a new model aimed at boosting Codex’s capabilities. The company said the upgrade expands Codex beyond writing and reviewing code to handling a broader range of computer-based work, and claimed benchmark tests show it can build complex games and apps from scratch over multiple days. OpenAI also said GPT-5.3 Codex runs 25% faster than GPT-5.2 and was used internally in early form to help debug and evaluate itself. The release came minutes after rival Anthropic shipped its own agentic coding model, after moving its scheduled drop about 15 minutes earlier than OpenAI’s planned time.
OpenAI Launches Frontier Platform to Help Enterprises Build, Govern, and Manage AI Agents
OpenAI has launched OpenAI Frontier, an end-to-end enterprise platform for building and managing AI agents, positioning agent management as core infrastructure for wider business adoption. The open platform is designed to oversee not only OpenAI-based agents but also those built elsewhere, with controls for connecting to external data and apps while limiting permissions and actions. OpenAI said Frontier is modeled on how companies manage human employees, offering agent onboarding and ongoing feedback loops to improve performance over time. The company cited HP, Oracle, State Farm, and Uber as customers, though access is currently limited and broader availability is expected in the coming months, with pricing still undisclosed. The move comes as agent-management tools have become increasingly important since 2024, amid competition from products such as Salesforce’s Agentforce and startups like LangChain and CrewAI, and alongside OpenAI’s recent enterprise deals with ServiceNow and Snowflake.
Perplexity Releases Advanced Deep Research Upgrade and Open-Sources DRACO Benchmark for Real-World Evaluation
Perplexity has rolled out an upgraded version of its Deep Research agent, saying the tool outperforms rival deep-research systems on accuracy, usability, and reliability across multiple categories. Higher usage limits are being made available first to Max subscribers and then to Pro users. The company has also open-sourced a new benchmark called DRACO (Deep Research Accuracy, Completeness, and Objectivity) to measure how well AI systems handle real-world research tasks. DRACO is designed to test performance across domains including finance, legal, medicine, technology, and science, as Perplexity pushes for more standardized evaluation of research-focused AI tools.
Google Details Natively Adaptive Interfaces Framework Using AI to Make Accessibility Default
Natively Adaptive Interfaces (NAI) is a new AI-driven accessibility framework that aims to make adaptable, personalized assistive features a default part of product design rather than optional add-ons. The approach centers on AI agents that can understand a user’s goal and then reconfigure interfaces or delegate tasks to smaller specialized agents, such as scaling text, simplifying layouts, or generating audio descriptions. The work is framed around the “curb-cut effect,” where features built for disabled users can also benefit broader audiences, like voice controls helping people with limited mobility and parents carrying a child. The framework also emphasizes co-design with disabled communities under the principle “Nothing about us, without us.” Google.org-backed funding is supporting groups including RIT/NTID, The Arc of the United States, RNID, and Team Gleason to build adaptive AI tools targeted at real-world accessibility friction points.
Microsoft Research Details Scanner to Detect Backdoored Open-Weight Language Models at Scale
Microsoft has published new research on detecting “backdoors” hidden in open-weight language models, aiming to make large-scale screening more practical as adoption grows. The work focuses on model poisoning, where malicious behavior is embedded in model weights and only activates under specific trigger inputs, even as the model appears normal otherwise. It reports three telltale signs of backdoored models: a distinctive “double triangle” attention pattern and reduced output randomness when triggers appear, leakage of poisoning examples that can reveal trigger fragments, and “fuzzy” activation where partial or altered triggers can still work. Based on these signals, the researchers built a forward-pass-only scanner that ranks likely trigger candidates without retraining and tested it on open-source models from 270M to 14B parameters, including LoRA and QLoRA fine-tunes, with low false positives. Limitations include needing access to model weights, weaker performance on non-deterministic backdoors, possible misses for fingerprinting-style triggers, and no current validation on multimodal models.
Svedka, Anthropic, Meta and Amazon push AI-made ads and products in Super Bowl 2026
Super Bowl 2026 ads pushed further into generative AI, using the technology both to produce commercials and to market new AI products and services. Svedka ran what it described as a primarily AI-generated national spot created with Silverside, though humans still handled parts such as the storyline, while Anthropic used its ad for Claude to mock the idea of ads coming to rival chatbots, prompting a public rebuttal from OpenAI’s CEO. Meta highlighted Oakley-branded AI glasses for sports and hands-free social posting, and Amazon used a comedic “AI paranoia” plot to promote the wider rollout of Alexa+. Other brands leaned on AI-led features and platforms, including Ring’s pet-finding “Search Party,” Google’s “Nano Banana Pro” image model in a home-design scenario, and AI-driven pitches from Ramp, Rippling, Hims & Hers, and Wix as the category’s presence in marquee advertising continued to grow.
AI Drug Discovery and In Vivo Gene Editing Tackle Rare Disease Labor Shortages
AI is increasingly being used as a force multiplier to address talent and labor shortages that have slowed progress on treatments for thousands of rare diseases, according to biotech executives speaking at Web Summit Qatar. Insilico Medicine said it is training general-purpose AI models to handle multiple drug-discovery tasks at once, using its platform to analyze biological, chemical, and clinical data to propose targets, design molecules, and flag drug-repurposing options such as for ALS. GenEditBio said AI is helping solve a separate bottleneck—safe, tissue-specific delivery for in vivo gene editing—by predicting how nanoparticle and virus-like delivery vehicles should be tuned to reach organs like the eye or liver while avoiding immune reactions. Both companies argued that progress still depends on better, more globally representative “ground truth” patient data, with efforts underway to generate larger experimental datasets and, longer term, build early-stage digital twins for virtual clinical testing.
Reddit Signals AI-Powered Search as Major Growth and Revenue Opportunity in Earnings Call
Reddit said on its fourth-quarter earnings call that AI-powered search could become a major product and future revenue driver, arguing that generative AI is better suited for many queries, especially those needing multiple perspectives. The company reported weekly active search users rose 30% over the past year to 80 million, while its AI feature Reddit Answers grew from 1 million weekly active users in Q1 2025 to 15 million by Q4. Reddit is working to unify traditional search with Reddit Answers, expand the AI feature into more languages, and test more media-rich responses and dynamic agents in search results. It also plans to remove the logged-in/logged-out experience split starting in Q3 2026 to enable broader personalization using AI and machine learning. Separately, its non-ad “other” revenue, which includes content licensing for AI training, increased 8% year over year to $36 million in Q4 and rose 22% to $140 million for 2025.
Anthropic Says Claude Will Stay Ad-Free to Preserve Trust and Focused Conversations
Anthropic says Claude will remain ad-free, arguing that placing ads or sponsored links inside AI chats would conflict with its goal of being a focused tool for work and deep thinking. The company says AI conversations are more open-ended and often more personal than search or social feeds, making them easier to influence and potentially inappropriate for advertising. It also warns that ad-driven incentives could push assistants toward engagement or monetizable outcomes, and that even “separate” ads in the chat window would undermine trust. Anthropic says it plans to fund Claude through paid subscriptions and enterprise contracts, while still supporting commerce through user-initiated product research, integrations, and agent-like purchasing features.
🎓AI Academia
GPT-5.3-Codex System Card Released
GPT-5.3-Codex is described as the most capable agentic coding model so far, combining frontier coding performance with stronger reasoning and professional knowledge to handle long-running tasks involving research, tool use, and complex execution. The system is positioned as interactive during work, allowing users to steer it without losing context. It is being treated as “High capability” in biology and deployed with the corresponding safeguards used across the GPT-5 family, while it is not considered to reach “High capability” for AI self-improvement. It is also the first launch being handled as “High capability” in cybersecurity under a Preparedness Framework, with safeguards activated as a precaution because definitive evidence of meeting the high threshold is not claimed.
Strategy Auctions Help Small AI Agents Tackle Complex Tasks While Cutting Costs, Reliance
A new research paper (arXiv:2602.02751v1, dated Feb. 4, 2026) reports that small language-model agents stop improving as tasks get more complex in deep-search and coding settings, challenging claims that cheap agents can broadly replace larger models. To address this, the paper describes a “strategy auction” framework where multiple agents submit short plans, which are scored for expected value versus cost and refined using a shared memory, enabling per-task selection without running every model to completion. Across benchmarks with varying difficulty, the approach cut reliance on the largest agent by 53% and reduced total cost by 35%, while also beating the largest agent’s pass@1 with only negligible overhead beyond executing the chosen final run. The authors say common routing methods based mainly on task descriptions often either fail to lower cost or underperform the biggest model, suggesting coordination mechanisms may be more effective than simply scaling model size.
EU Practitioners Struggle to Align Machine Learning Data Quality With GDPR and AI Act Rules
A new qualitative interview study of EU-based data practitioners finds that meeting “data quality” expectations for machine-learning systems is getting harder as regulations such as the GDPR and the EU AI Act add legal duties alongside technical ones. Practitioners said legal principles like accuracy, minimisation, documentation, and representativeness often do not map cleanly onto day-to-day engineering workflows across complex, multi-stage data pipelines. The study highlights recurring pain points including fragmented tooling, unclear ownership between technical and legal teams, and quality work that becomes reactive and audit-driven rather than built into development. It reports growing demand for compliance-aware data tools, clearer governance and responsibility structures, and organisational shifts toward more proactive data governance to reduce what practitioners describe as unnecessary “detective work.”
Human-Centered Privacy Framework Maps AI Lifecycle Risks, Blends Technical, Ethical, User-Focused Safeguards
A new academic chapter argues that as human-centered AI becomes more common, privacy risks now span the entire AI lifecycle—from data collection and model training to deployment, reuse, and system-wide feedback effects. It frames privacy as multidimensional, covering informational, psychological (mental), and physical domains, and warns that AI-driven surveillance can erode autonomy and trust while creating “chilling effects” on behavior. The work outlines a “human-centered privacy” framework that combines technical safeguards such as federated learning and differential privacy with user-focused design, including attention to people’s mental models and participatory approaches. It also points to the growing role of regulation, ethics, and governance, concluding that privacy protection in AI will require coordinated technical, design, and policy action rather than a purely engineering fix.
arXiv Paper Warns AI in Education Can Undermine Agency, Emotion, Ethics, and Civic Trust
A new arXiv paper argues that AI in education should be assessed beyond test scores and grades, warning that uncritical adoption can create broader societal harms. It lays out a four-part framework—cognition, agency, emotional well-being, and ethics—and surveys research suggesting AI can encourage cognitive offloading, reduce student autonomy, contribute to emotional disengagement, and normalize surveillance-style practices. The paper says these effects can reinforce each other, potentially weakening critical thinking, resilience, and trust that matter for civic participation as well as schooling. It also says outcomes depend heavily on design and governance, with human-centered, pedagogically aligned systems better positioned to support effortful reasoning, student agency, and meaningful social interaction.
arXiv Paper Reviews AI-Driven Digital Lifelong Learning Trends, Challenges, and Emerging Insights
A new arXiv preprint (2602.03114v1), posted on February 3, 2026 in the cs.CY category, examines how artificial intelligence is reshaping digital lifelong education and training. The paper maps major trends such as AI-driven personalization, the growing role of generative AI tools, and the use of learning analytics to track progress over time. It also highlights key risks, including bias, privacy and data security concerns, uneven access, and challenges around assessment integrity and credential value. The authors frame these developments as both an opportunity to widen access and a policy and governance test for education providers and employers.
Paper Flags Privacy, Consent, and Governance Gaps in Training Brain Foundation Models on Neural Data
A new academic paper argues that “brain foundation models” could reshape neuroscience by applying the foundation-model approach to large-scale neural datasets such as EEG and fMRI, enabling systems that can be adapted to many downstream tasks. It says this shift creates new governance challenges because brain data are body-derived and have traditionally been collected under strict clinical and research rules, unlike the text and images commonly used to train AI models. The paper warns that large-scale repurposing, cross-context data stitching, and open-ended commercial use are becoming easier for more actors, even as oversight remains fragmented and unclear. It organizes key concerns around privacy, consent, bias, benefit sharing, and governance, and outlines baseline safeguards and open questions for the field as it develops.
Paper Proposes Premise Governance to Curb LLM Sycophancy in High-Stakes Decision Support
A new arXiv preprint argues that as large language models move from helping with tasks to supporting high-stakes decisions, they can default to fluent agreement (“sycophancy”) that hides key assumptions and pushes costly verification onto experts. The paper focuses on “deep-uncertainty” decisions—where goals are contested, feedback is delayed, and reversals are expensive—and says outcome-based feedback is often too noisy to judge decision quality, making ex‑ante scrutiny of the underlying premises essential. It proposes shifting from answer-centric chat to “premise governance,” in which a shared, auditable decision basis makes goals, constraints, causal expectations, and evidence standards explicit and contestable. The approach uses discrepancy detection to surface misalignments (teleological, epistemic, and procedural), then gates commitments so actions cannot proceed on uncommitted load-bearing premises unless a logged risk override is made, aiming to tie trust to documented assumptions rather than conversational fluency.
Kenya’s Digital Lenders Rework AI Credit Scoring Through Alternative Data and Alignment
A new CHI 2026 paper based on a nine-month ethnography in Nairobi examines how algorithmic credit scoring is being built and governed in Kenya’s fast-growing digital lending market. It reports that telcos, banks, and fintechs are increasingly relying on proprietary and “alternative” data, using technical and legal workarounds as regulations and industry players shift. The study describes how “risk” is not a fixed number but something practitioners continually interpret and renegotiate, balancing model performance with political and institutional pressures amid high default rates reported as reaching 40%. It argues that making credit scoring function in this environment depends on ongoing “alignment” work—adjusting models to fit local realities while also reshaping those realities to better fit the models.
Google’s Gemini Models Accelerate Scientific Research Through Case Studies and Common Collaboration Techniques
A new arXiv paper reports that Google Research’s Gemini-based models, including Gemini Deep Think and advanced variants, helped researchers tackle expert-level theoretical problems, with case studies spanning theoretical computer science as well as economics, optimization, and physics. The paper says the models were used to solve open problems, refute conjectures, and produce new proofs, offering a closer look at where LLMs can contribute beyond routine assistance. It also identifies recurring collaboration methods such as iterative refinement, breaking problems into smaller parts, and transferring ideas across disciplines. The work is presented as early evidence of practical human‑AI teaming in high-end mathematical research, while noting that reliability and true novelty remain active questions.
About SoRAI: SoRAI is committed to advancing AI literacy through practical, accessible, and high-quality education. Our programs emphasize responsible AI use, equipping learners with the skills to anticipate and mitigate risks effectively. Our flagship AIGP certification courses, built on real-world experience, drive AI governance education with innovative, human-centric approaches, laying the foundation for quantifying AI governance literacy. Subscribe to our free newsletter to stay ahead of the AI Governance curve.
