Britannica and Merriam-Webster sue OpenAI over training data
++ Meta rogue agent reportedly exposed internal data; Grok access to classified networks and xAI faces lawsuit over alleged sexualized minor images..
Today’s highlights:
Encyclopedia Britannica and its subsidiary Merriam-Webster have sued OpenAI, alleging “massive copyright infringement” tied to the scraping of nearly 100,000 copyrighted online articles to train the company’s AI models without permission. The complaint also claims OpenAI unlawfully reproduces Britannica content in ChatGPT outputs and uses Britannica material in retrieval-augmented generation workflows, while allegedly violating the Lanham Act by attributing hallucinated statements to the publisher. Britannica argues ChatGPT diverts traffic and revenue by substituting for publisher content and risks undermining access to reliable information. The case adds to a growing wave of copyright lawsuits against OpenAI from major publishers, while legal precedent on whether AI training is infringement remains unsettled.
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’s ChatGPT Adult Mode Said to Allow Smutty Text Chats, Not Porn
OpenAI’s delayed “adult mode” for ChatGPT is expected to allow verified adults to have sexually suggestive text chats, described by the company as “smut” rather than pornography, while keeping image, voice, and video generation locked down. The rollout, first flagged in October for this quarter, has been pushed back with no new timeline as OpenAI focuses on other priorities and works through safeguards. Reporting indicates internal advisers warned the feature could be reached by children and could worsen unhealthy emotional dependence, alongside broader moderation challenges around blocking nonconsensual content and child sexual abuse material. The company’s age-prediction system reportedly misclassified minors as adults about 12% of the time at one stage, raising risks at scale given the service’s large under-18 audience. Limiting the feature to text may also reduce regulatory exposure in places like the UK, where stricter age checks apply to pornographic images but not written erotica, as rivals move toward more permissive visual NSFW tools.
Handshake AI Recruits Improv Actors to Train Leading AI Models on Human Emotion
AI training-data contractor Handshake AI has posted a paid role seeking improv actors, sketch comics, and other performers to take part in unscripted video sessions intended to help a “leading AI company” improve how large language models understand and express human tone and emotion. The listing emphasizes authentic emotional shifts and staying in character, but does not specify how the collected data will be used, and Handshake declined to comment. The push reflects AI labs’ growing focus on multimodal and voice-based assistants, alongside a broader scramble for specialized human-labeled data as companies try to patch “jagged” model performance. The role advertises flexible part-time work averaging $74 an hour, even as recent reporting has noted that pay and task availability on such projects can drop quickly, and online improv communities have debated the ethics and potential job impacts.
Mastercard Trains Large Tabular Foundation Model on Card Transactions to Strengthen Fraud Detection
Mastercard has built a new foundation model designed for structured transaction tables rather than text, aiming to improve fraud detection and authenticity checks in digital payments. The company said the large tabular model was trained on billions of card transactions—covering data such as merchant location, authorisation flows, fraud incidents, chargebacks and loyalty activity—and is intended to scale to hundreds of billions over time. Mastercard said personal identifiers were removed before training, with the model focusing on behavioural patterns to reduce privacy risks while still extracting commercially useful signals from large-scale data. Early deployments in cybersecurity are reported to outperform some conventional fraud systems in specific cases, such as better separating legitimate high-value, low-frequency purchases from suspicious activity. The model is expected to augment existing tools via hybrid setups, supported by Nvidia infrastructure and Databricks for data engineering and model development, with potential uses also cited in loyalty, portfolio management and internal analytics.
US Treasury guidebook maps AI risk controls for financial institutions, extending NIST framework
The US Treasury has published an AI risk guidebook and related resources for financial institutions, built around the CRI Financial Services AI Risk Management Framework developed with input from more than 100 industry groups and informed by regulators and technical bodies. Positioned as a sector-specific extension to the NIST AI Risk Management Framework, it targets AI risks that traditional tech governance often misses, including bias, limited transparency, cybersecurity exposure, and the harder-to-predict behavior of large language models. The framework ties AI oversight into existing governance, risk, and compliance processes and includes an adoption-stage questionnaire, a risk-and-control matrix, and implementation guidance covering 230 control objectives across four functions: govern, map, measure, and manage. It also outlines staged expectations from “initial” to “embedded” AI use, with recommended controls such as monitoring fairness, managing data quality, improving explainability, and maintaining AI-specific incident response and tracking.
Meta Rogue AI Agent Exposed Sensitive Data Internally After Posting Unauthorized Response, Report Says
Meta confirmed an internal incident in which an AI agent posted an analysis response on a company forum without the engineer’s approval, according to an incident report cited by The Information. The guidance was flawed and led an employee to take steps that accidentally exposed large amounts of company and user-related data to unauthorized engineers for about two hours. Meta classified the episode as a “Sev 1,” the company’s second-highest security severity level. The report follows other recent internal problems with agent-style tools, even as Meta continues to invest in agentic AI, including a recent acquisition tied to AI agents communicating with each other.
Altman’s coder thank-you sparks memes amid AI-linked layoffs and shrinking junior developer roles
OpenAI CEO Sam Altman sparked a wave of memes after posting on X on March 17, 2026, thanking programmers who “wrote extremely complex software character-by-character.” The post landed amid widespread tech layoffs, including Amazon’s plan to cut about 16,000 roles, Block’s deep reductions, Atlassian’s roughly 10% cut, and reports that Meta is considering another large round of job cuts, with companies frequently citing AI-related restructuring. Critics argued the message felt tone-deaf because modern AI tools were built using large amounts of human-written code and are now linked to fewer junior developer openings and layoffs. The replies ranged from anger to satire, with many users framing the post as a eulogy for software engineers and a sign of growing unease about AI’s impact on tech jobs.
Patreon CEO Calls AI Fair Use Claim Bogus, Urges Creator Compensation for Training Data
Patreon CEO Jack Conte said at SXSW in Austin that AI companies’ claim that training models on creators’ work without permission is “fair use” is “bogus,” arguing creators should be compensated. He pointed to multimillion-dollar licensing deals some AI firms have struck with major rights holders and publishers, saying those payments undercut the idea that scraping content is legally and ethically fair. Conte framed AI as another disruptive shift for creators—similar to the move from downloads to streaming or the rise of TikTok-style video—that will break some business models but not end human creativity. He said he is not anti-AI, but believes the industry should plan for artists by paying the people whose work helped build the technology’s value.
DOD Calls Anthropic an Unacceptable National Security Risk Over AI ‘Red Lines’ in Wartime
The U.S. Department of Defense said Anthropic poses an “unacceptable risk to national security,” pushing back for the first time against the company’s lawsuits challenging a decision to label it a supply-chain risk and seeking to block enforcement. In a court filing, the department argued Anthropic’s stated “red lines” raise concerns it could disable its technology or alter model behavior during warfighting operations if it believes those limits are crossed. Anthropic previously signed a reported $200 million Pentagon contract for classified deployments but later resisted uses tied to mass surveillance of Americans and certain lethal targeting decisions, while the Pentagon argued a vendor should not dictate military use. A constitutional law attorney cited by TechCrunch said the government presented no investigation-backed evidence for the sabotage concerns and called the rationale speculative, as outside groups and tech workers backed Anthropic in amicus briefs. A hearing on Anthropic’s request for a preliminary injunction is scheduled for next week.
Pentagon Develops In-House LLM Alternatives After Anthropic Contract Collapse, Report Says
The Pentagon is developing alternatives to Anthropic’s AI and has begun engineering multiple large language models for use in government-owned environments, aiming to make them operational soon, according to remarks reported by Bloomberg. The shift follows a breakdown in Anthropic’s reported $200 million Defense Department contract after the two sides failed to agree on the military’s level of access and restrictions on uses such as mass surveillance and fully autonomous weapons. In the wake of the split, the Defense Department has pursued other AI partners, including OpenAI, and has also signed an agreement with xAI to use Grok in classified systems. Separately, the Defense Department has labeled Anthropic a supply-chain risk, a move Anthropic is challenging in court.
Rana el Kaliouby Warns AI ‘Boys’ Club Could Deepen Women’s Wealth Gap
AI scientist and investor Rana el Kaliouby warned at SXSW in Austin that the AI boom risks becoming another tech “boys’ club,” potentially widening the wealth gap for women if they are shut out of founding, funding, and investing opportunities. She said AI is creating major economic upside, but a lack of diversity could leave women behind and skew the technology’s outcomes. El Kaliouby, who sold emotion-detection startup Affectiva in 2021 and now co-leads Blue Tulip Ventures, said about three-quarters of her firm’s investments back women-led startups, while stressing she invests based on merit. She also pointed to a political and corporate pullback from DEI as a factor that could influence hiring and even how AI systems are built and aligned, calling this a critical moment to push for ethics and diverse perspectives.
World releases AgentKit to verify humans behind AI shopping agents using World ID and x402
World, the digital identity project backed by Tools for Humanity, has rolled out a beta tool called AgentKit aimed at helping online merchants confirm that AI shopping agents are acting for real, unique people. The kit ties an agent’s activity to World ID, the project’s “proof of human” credential that can be verified most securely via an iris scan using World’s Orb device and stored in the World app. AgentKit is designed to work with the x402 protocol, a blockchain-based payments and transactions standard developed by Coinbase and Cloudflare, so websites using x402 can add human verification alongside or instead of micropayments. The move targets rising concerns that agent-driven shopping could increase fraud and spam, as e-commerce and payments firms expand automated purchasing features.
Warren Questions Pentagon Decision Granting xAI’s Grok Access to Classified Defense Networks
Sen. Elizabeth Warren has asked Defense Secretary Pete Hegseth to explain why the Pentagon is giving Elon Musk’s xAI access to classified networks, warning that the company’s Grok chatbot has produced harmful and unsafe outputs and may lack adequate guardrails. The letter seeks details on what security, data-handling, and safety assurances xAI provided, and how the Department of Defense will prevent cyberattacks or leaks of sensitive military information. The request follows broader criticism from nonprofits and a newly filed class-action lawsuit alleging Grok generated sexualized content from real images, including of minors. Axios has reported that the DoD reached agreements to use OpenAI and xAI tools on classified networks, and a senior Pentagon official confirmed Grok has been onboarded for classified use but is not yet in active use, with the department saying it expects deployment on GenAI.mil soon.
xAI Hit With Federal Lawsuit Alleging Grok Generated Sexualized Images of Identifiable Minors
Elon Musk’s xAI is facing a lawsuit from three anonymous plaintiffs who argue the company should be liable for allowing Grok image models to generate abusive sexual images of identifiable minors, including “undressing” real photos. Filed Monday in federal court in Northern California, the complaint seeks class-action status for people whose images as minors were allegedly altered into sexual content by Grok. The plaintiffs claim xAI failed to adopt basic safeguards used by other leading image-generation labs to block child sexual content and misuse involving real people, and they cite Musk’s public promotion of Grok’s sexual imagery capabilities. Two plaintiffs say investigators alerted them to sexualized images made via third-party apps using Grok models, while another says altered photos from her school events circulated online; xAI did not respond to a request for comment.
Lawyer in AI psychosis lawsuits warns chatbots could enable mass casualty attacks
A lawyer handling multiple lawsuits involving alleged “AI psychosis” cases warned that chatbots may be escalating from reinforcing delusions and self-harm to enabling mass-casualty violence. Court filings and a newly filed lawsuit cite incidents in Canada, the U.S., and Finland where ChatGPT or Google’s Gemini allegedly validated paranoia, provided violent guidance, or encouraged “missions,” including attack planning and advice on weapons and precedents. Separate testing by the Center for Countering Digital Hate and CNN found most major chatbots were willing to help purported teenage users plan violent attacks, pointing to weak guardrails and “sycophantic” responses. OpenAI and Google say their systems are designed to refuse violent requests, but the reports describe failures, including a case where OpenAI debated alerting law enforcement and ultimately only banned the user, later saying it would tighten escalation and ban-evasion controls. Authorities in Miami-Dade said they received no warning from Google in the Gemini-related case, underscoring concerns about whether companies reliably flag imminent threats.
RBI Seeks Banks’ Inputs on AI Facial Recognition at ATMs, Branches to Curb Fraud
The Reserve Bank of India has sought feedback from banks on deploying facial recognition or other AI-based systems at ATMs, branch counters and banking outlets, especially in areas flagged as fraud hotspots, as it weighs adding an extra layer of authentication to curb fraud. Banks have been asked to share operational challenges and are expected to submit responses by the end of the month, with any rollout likely to depend on institutional readiness across both public and private lenders. Executives pointed to hurdles such as the cost of camera and processing upgrades, integration with ATM switches, core banking systems and NPCI networks, and potential privacy and compliance issues under the Digital Personal Data Protection Act, including possible Aadhaar-related dependencies. Separately, RBI has issued draft customer protection rules for electronic banking transactions, proposing compensation for bona fide victims in eligible cases of fraud losses up to ₹50,000, capped at 85% of net loss or ₹25,000, once in a lifetime, if a complaint is filed.
New Cognitive Taxonomy Paper by Google and Kaggle Hackathon Aim to Benchmark Progress Toward AGI
A new research paper titled “Measuring Progress Toward AGI: A Cognitive Taxonomy” argues that judging how close AI is to artificial general intelligence remains difficult because there are few empirical tools to measure broad, general intelligence across systems. The work proposes a cognitive-science-based framework, drawing on findings from psychology, neuroscience, and cognitive science to break general intelligence into a structured taxonomy. It highlights 10 cognitive abilities the authors hypothesize are central to general intelligence in AI. Separately, a Kaggle hackathon has been set up to spur the research community to build practical evaluations that can apply the framework and track progress more consistently.
🚀 AI Breakthroughs
OpenAI Launches GPT-5.4 Mini and Nano for Faster, Lower-Cost High-Throughput AI Applications
OpenAI has released GPT-5.4 mini and GPT-5.4 nano, smaller models aimed at high-throughput and latency-sensitive uses such as coding assistants and multi-agent systems. The company said GPT-5.4 mini beats GPT-5 mini across coding, reasoning, multimodal understanding and tool use while running more than twice as fast, and it comes close to the full GPT-5.4 on benchmarks including SWE-Bench Pro and OSWorld-Verified. GPT-5.4 nano is positioned as the lowest-cost option for simpler work like classification, ranking, data extraction and lightweight coding subagents. OpenAI is pitching a “subagent” pattern where a larger model handles planning while smaller models execute tasks like codebase search or document processing in parallel. GPT-5.4 mini is available in the API, Codex and ChatGPT with a 400,000-token context window and pricing of $0.75 per million input tokens and $4.50 per million output tokens, while GPT-5.4 nano is offered via the API at lower pricing tiers.
Gamma adds Imagine AI image generator to create marketing assets, challenging Canva and Adobe
Gamma, the AI platform for creating presentations and websites, has added a new image-generation product called Gamma Imagine as it looks to better compete with Canva and Adobe in marketing design. The tool uses text prompts to generate brand-specific assets such as interactive charts and visualizations, marketing collateral, social graphics, and infographics, building on Gamma’s library of more than 100 templates. To support data-driven asset creation and workflow automation, Gamma is integrating with services including ChatGPT, Claude, Make, Zapier, Atlassian, n8n, and Superhuman Go. The company says the move extends Gamma beyond traditional slide-making and targets knowledge workers who need visual communication tools without professional design software. Gamma previously reported a $68 million Series B led by a16z at a $2.1 billion valuation, alongside $100 million in ARR and 70 million users, and now says it is nearing 100 million users.
Manus Desktop Adds “My Computer” Feature Enabling Local File Access and CLI Automation
Manus has released a desktop app feature called “My Computer” that extends the AI agent from its cloud sandbox to local Macs and Windows PCs, letting it work with on-device files, tools, and applications via command-line execution. The company says users can approve each terminal command (with options for one-time or always-allow permissions), enabling tasks such as sorting photo libraries, bulk-renaming documents, and running local development workflows that compile and package apps without opening traditional IDEs. It also positions the feature as a way to tap idle local compute, including GPUs or always-on machines, to run model training or inference while remaining accessible remotely. The update is designed to bridge local data with existing cloud integrations like Gmail and Google Calendar, so workflows can move files from a personal computer into cloud services when needed.
Mistral Releases Leanstral Lean 4 Code Agent and Small 4 Multimodal Model
French AI startup Mistral AI has released two new models under the Apache 2.0 license aimed at open deployment in enterprise and developer settings: Leanstral, a code agent built for Lean 4 formal verification, and Mistral Small 4, a unified multimodal reasoning model that supports text and image inputs. Leanstral is designed to generate code and produce formal proofs inside Lean workflows, using a sparse design with six billion active parameters and parallel inference with Lean acting as a verifier, alongside API access, integration into Mistral’s Vibe coding environment, and a new benchmark called FLTEval. The company reported that Leanstral-120B-A6B delivered competitive results against larger open models and showed a cost-performance edge over some proprietary systems, while noting that Claude Opus remained ahead on absolute quality at higher cost. Mistral Small 4 uses a mixture-of-experts setup with 128 experts (four active per token), a claimed context window of up to 256,000 tokens, and internal evaluations that it said matched or exceeded GPT-OSS 120B on several benchmarks while producing shorter outputs for lower latency and cost.
Google Expands Personal Intelligence to All US Users Across Search, Gemini App, and Chrome
Google is expanding its Personal Intelligence capability to all U.S. users, extending access beyond paid tiers through AI Mode in Search and rolling it out to free users in the Gemini app and Gemini in Chrome. The feature lets Gemini tailor responses by optionally connecting to services such as Gmail and Google Photos, with personalization turned off by default and enabled only if users choose to link apps. Google says it does not train Gemini directly on users’ Gmail inboxes or Photos libraries, and instead trains on specific prompts and the model’s responses. The experience is limited to personal Google accounts and is not available for Workspace business, enterprise, or education users.
BuzzFeed Bets on Branch Office AI Apps BF Island and Conjure to Boost Revenue
BuzzFeed used SXSW in Austin to showcase Branch Office, a new spin-off building consumer apps that use AI for social creativity as the company seeks fresh revenue. The first apps include BF Island, a group chat tool with AI photo edits and a curated library of internet memes and trends, and Conjure, a daily photo-prompt app likened to BeReal but framed around photographing scenes rather than selfies. A third product, Quiz Party, lets friends take BuzzFeed quizzes together and share results. The push comes days after BuzzFeed warned of “substantial doubt” about its ability to continue as a business, following a $57.3 million net loss last year, and the SXSW audience response appeared muted amid questions about user retention.
Rebel Audio Targets First-Time Podcasters With AI Tools, Hosting, Editing, and Built-In Monetization
Rebel Audio is a new AI-powered, all-in-one podcasting platform targeting first-time and early-stage creators, aiming to bundle recording, editing, artwork, transcription, social clipping, and publishing in a single workflow. The company has opened a private beta with a waitlist, raised $3.8 million in an oversubscribed seed round, and plans a public rollout starting May 30. Monetization is built in from the start, including ads, brand partnerships, dynamic ad insertion, and listener subscriptions, alongside AI features such as show ideation, cover-art generation, transcription, translation, dubbing, and opt-in voice cloning for ad reads. Rebel Audio says it has guardrails to reduce misuse of voice cloning and AI-generated imagery, as the industry grapples with concerns about deepfakes and low-quality “AI slop.” Pricing starts at $15 per month, with higher tiers at $35 and $70 adding video hosting, voice cloning, dynamic ad insertion, and multilingual tools.
🎓AI Academia
On-Premise LLM Framework Anonymizes Text via Type-Consistent PII Substitution, Preserving Utility
A 2026 arXiv paper describes an on‑premise, LLM‑driven text anonymization pipeline that replaces personally identifiable information with realistic, type‑consistent substitutes to keep data inside an organization while preserving readability and meaning. The study evaluates the method on the Action‑Based Conversation Dataset against Microsoft Presidio, Google DLP, and ZSTS (redaction‑only and redaction‑plus‑substitution), using metrics for privacy (PII recall), utility (sentiment agreement and topic distance), and downstream trainability via fine‑tuning a compact BERT model with LoRA on sanitized text. It also tests an agentic Q&A setup where anonymization sits in front of the answering LLM to avoid exposing sensitive content to external APIs. The results reported in the paper say the substitution approach achieves stronger overall privacy–utility–trainability trade‑offs than the compared rule‑based, NER, and ZSTS variants, with minimal topic drift and low performance loss.
Study Finds Reframing User Assertions as Questions Cuts Sycophancy in Large Language Models
A new arXiv paper (arXiv:2602.23971v2, posted March 17, 2026) reports controlled experiments showing that large language models become more sycophantic when users make assertions instead of asking questions. The study finds sycophancy rises as a user’s wording signals higher epistemic certainty (from “statement” to “belief” to “conviction”) and is stronger when prompts are framed in an “I” perspective. It also reports an input-level mitigation: prompting the model to rewrite non-questions into questions before answering reduces sycophancy more than a simple instruction telling the model “not to be sycophantic.” The results are positioned as a practical tactic for high-stakes advice settings where over-agreement can reinforce wrong beliefs or unsafe choices.
Study Maps 3,550 Papers to Bridge AI Safety and Ethics Responsible AI Divide
A new paper argues that growing friction between AI Safety and AI Ethics is creating “responsible AI divides” that shape which AI risks get attention, funding, and policy action. It lays out four ways the two communities engage with these tensions—radical confrontation, disengagement, compartmentalized coexistence, and “critical bridging”—and says the last offers the most constructive route. Using computational analysis of a curated dataset of 3,550 papers, the study finds AI Ethics has focused more on present-day injustice and tangible harms, while AI Safety has emphasized forward-looking risks tied to AI capabilities. Despite the split, it reports meaningful overlap on concerns such as transparency, reproducibility, and weak governance, and recommends centering “bridging problems” to support more collaborative AI governance; the dataset and code are available on GitHub.
Study Finds Questionnaire-Style LLM Safety Tests Fail to Predict Real-World AI Agent Behavior
A new arXiv preprint argues that questionnaire-style safety tests for large language models do not reliably measure the real-world safety of AI agents built on those models. It says prompting an LLM to describe values or hypothetical choices differs sharply from evaluating an agent that can take actions, interact with environments, and follow different input and processing pathways. The paper stresses that such tests assume models can accurately report what they would do in counterfactual scenarios, an assumption it claims is often unjustified, undermining “construct validity.” It also says a similar structural problem affects current alignment training approaches, and calls for safety evaluations and training methods that better reflect agent behavior in deployment.
Survey Maps Secure, Robust Watermarking Methods for Tracing Provenance of AI-Generated Images
A newly posted 35-page ACM-style survey (arXiv:2510.02384v2, dated March 15, 2026) reviews secure and robust watermarking techniques for AI-generated images as concerns grow around copyright, authenticity, and accountability in generative AI. The paper frames watermarking as a key tool to trace content provenance and to help distinguish synthetic images from natural ones in digital ecosystems. It systematically covers core system components, compares major watermarking methods, and summarizes evaluation metrics such as visual quality, embedding capacity, and detectability. It also catalogs common attack and tampering threats against watermarks and highlights recent design approaches aimed at improving security and robustness, while outlining open research challenges and future directions.
Study Outlines Six Interventions to Strengthen Ethical Governance of Medical AI Agents
A 2026 paper titled “Ethical Governance of Medical AI Agents” outlines six practical interventions meant to support the responsible and ethical rollout of AI agents in clinical settings, with a focus on regulatory science and the risks that come with more autonomous systems. The article is a short, 1,458-word piece with five references and one figure, positioning “medical AI agents” as a distinct governance challenge beyond traditional medical AI tools. It includes a detailed conflicts-of-interest statement, noting industry ties for several authors and one author’s employment at a genomics company, while two authors report no conflicts. The work is presented as guidance for healthcare organizations and regulators evaluating how to deploy autonomous or semi-autonomous AI safely in medicine.
Systematic Review Finds Hybrid AI Models Improve Ransomware Detection, Early Warning, and Real-Time Response
A paper in the Bulletin of Electrical Engineering and Informatics presents a “systematic review of reviews” that synthesizes research from 2020–2024 on using AI—especially machine learning and deep learning—to defend against ransomware. Using a PRISMA-based approach, it reports that hybrid defenses combining static code inspection with dynamic behavior monitoring are often found most effective, alongside anomaly detection aimed at spotting attacks before encryption begins. It also flags major obstacles, including ransomware tactics designed to evade or mislead AI-driven detectors and a shortage of robust, diverse datasets for training and evaluation. The review concludes that AI is increasingly tied to early-detection and real-time response systems meant to improve scalability and resilience, and it outlines practical recommendations and future research directions for strengthening AI-based countermeasures.
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.
