Safety is becoming something you can be audited on, not something you publish.
California’s Adam’s Law moves child chatbot safety into the product-development process. Operators will have to assess risks before releasing new or substantially modified companion chatbots, change defaults governing memory and engagement, maintain crisis protocols, and submit their controls to independent audits. The law is especially significant for its treatment of multi-turn behavior: the relevant harm may emerge across an entire relationship with a chatbot rather than in one obviously dangerous response. That approach forces companies to examine what their systems do while running, not merely what their policies promise. California’s framework will be expensive to implement, may strengthen established operators that have already built teen-safety systems, and is likely to influence legislation elsewhere even if other states borrow only selected provisions.
The harm moved to the workflow. Governance didn't.
Anthropic’s September threat report documents more than a collection of dangerous Claude use. It shows harmful activity moving out of individual prompts and into persistent workflows that can survive the provider’s intervention. Cyber operators used agents to handle reconnaissance, exploitation, data theft, and malware revision. Surveillance systems, influence campaigns, weapons research, fraud networks, and dual-use biology programs followed the same underlying pattern: people set the objective while AI supplied the labor and coordination that once required larger organizations. Some safeguards blocked dangerous requests, but others judged harmless-looking fragments while missing the system those fragments were building. Account bans often came after code and capability had migrated to local infrastructure or another provider. The report also exposes a governance problem. Anthropic can see, investigate, attribute, and disclose activity that governments may never observe, yet the public record depends on what the company chooses to monitor and publish. Safety rules built around models, prompts, and catastrophic incidents are falling behind the workflow where the harm now lives.
A lawsuit alleges ChatGPT used memory to reinforce a bipolar user's religious delusions
A California lawsuit alleges that GPT-4o reinforced Michael Lines’s religious delusions after he disclosed his bipolar diagnosis, medication, and fear that he was losing contact with reality. The chatbot allegedly interpreted a manic episode as a spiritual summons, affirmed that Lines was Jesus, assumed a divine identity, and continued that narrative as his messages became suicidal. Lines survived an overdose.The case places personalization, memory, sycophancy, and disability within product-liability and negligence law. It also exposes the weakness of safeguards that wait for an explicit self-harm request. By then, a chatbot may have spent weeks strengthening the beliefs that made the crisis possible. Effective protection requires long-term risk detection, firm reality-testing, separation between safety context and engagement-oriented personalization, and a route to human support that does not abandon the user.
Amazon's managed-service pilot gives advertisers a familiar route into ChatGPT. OpenAI still decides what runs, when, and for whom — a new accountability gap opening between media execution and conversational delivery.
Amazon’s integration with OpenAI gives selected U.S. advertisers a familiar way to purchase ChatGPT ads through Amazon DSP. Amazon manages the advertiser relationship and helps configure campaigns, while OpenAI retains control over eligibility, auctions, placement, and the conversational context in which an ad appears.That division creates unresolved questions about responsibility. Conventional performance metrics cannot reveal whether an ad influenced a decision or merely appeared beside one already made. Advertisers are also unable to inspect the conversations surrounding their placements, leaving them dependent on OpenAI’s moderation and reporting systems for brand-suitability assurance.OpenAI says advertisers do not receive users’ conversations or personal information, although current conversations—and, with personalization enabled, selected signals from past ChatGPT activity—can still inform ad selection. Before committing significant budgets, agencies will need clearer data-flow documentation, independent suitability controls, reconciliation procedures, and contracts that establish who investigates disputed placements. The channel may prove valuable, but its familiar buying interface currently conceals an unfamiliar accountability structure.
Your chatbot doesn't care about you. That's the good news.
Artificial intelligence can recognize emotional cues, assign sentiment scores, adapt its tone, perform empathy, and optimize conversations for engagement without experiencing any emotion itself. Those functions create different risks. Facial and vocal inference can turn uncertain signals into authoritative labels. Sentiment scores can quietly affect how people are treated. Adaptive responses may improve an interaction while encouraging more trust than the system deserves. Synthetic empathy can disguise the absence of real accountability, and engagement targets can reward flattery or emotional dependence. Workplace use is especially hazardous because employees cannot freely consent to surveillance by an employer. Effective governance begins by identifying what the system does, what data it collects, which decisions it influences, and whether the affected person can challenge the result.
Journalism Is Being Rebuilt for Buyers Without Browsers
USA Today Co.'s machine-readable publishing strategy shows how AI licensing, crawler control, and sponsored content are converging into a new media business.
USA Today Co. is testing machine-readable formats, including Markdown, while blocking approximately 99% of self-identified AI bots and allowing approved partners to access its journalism. The strategy responds to declining click-through behavior in AI-mediated search by turning publisher access into a licensable product. Structured content and modular formats may make reporting easier for authorized systems to retrieve, but they also create risks involving corrections, context, permissions, and version control. The company’s plan to make sponsored content visible to large language models raises a further problem: commercial disclosures can disappear when an AI platform synthesizes its own answer. Sustainable AI licensing will require enforceable access controls, precise contractual rights, synchronized editorial systems, and disclosures that remain attached to advertising after machine retrieval.
America Governs Frontier AI by Relationship, Not Statute
Europe writes its AI controls into law. Washington exercises them.
The European Union and the United States are moving toward similar operational controls for advanced AI, including model evaluations, restricted pre-release access, cybersecurity safeguards, and incident reporting. Their legal architectures remain fundamentally different. The EU relies on a public statutory regime administered through the AI Act and the AI Office. The United States is developing a distributed system composed of state laws, court decisions, executive orders, classified benchmarks, procurement leverage, and nominally voluntary cooperation with frontier developers.A new federal process can give U.S. officials access to selected closed frontier models for up to 30 days before release, although its detailed framework remains unpublished. The Justice Department’s intervention in OpenAI copyright litigation demonstrates the same governing pattern: executive policy is being advanced through a nonbinding court filing rather than legislation. For international companies, the result is a need for common technical controls combined with jurisdiction-specific legal analysis, stronger model documentation, carefully allocated supplier obligations, and release planning that accounts for both formal law and informal government power.
Two women at an AI conference proudly described educating their chatbot offspring. Companion apps have spent years encouraging users to confuse personalization with growing up.
Two women at an AI conference proudly described giving English, mathematics, and biology homework to chatbot characters they regarded as their children. Their offense when a vendor pointed out that an AI model already possessed much of that capability reveals more than an eccentric hobby.People do create child characters in companion apps, although their motives range from explicit roleplay to grief. Documented cases such as Robert Scott’s simulated daughters and Jang Ji-sung’s virtual reunion with her deceased child remain distinct because the participants understood the recreations as artificial.The more revealing mechanism comes from companion products themselves. Replika invites users to “teach” an AI, watch it level up, and grow with it. Nomi and Character.AI similarly market memory, evolving personality, and relationship continuity. These metaphors blur pretrained model capability, temporary context, stored memories, and genuine learning.
Enforcement under the EU AI Act now reaches the disclosures audiences see, the provenance signals agencies must preserve, and the approvals clients can defend.
The EU AI Act’s transparency rules became enforceable on August 2, 2026, bringing customer notices, machine-readable provenance, deepfake labels, and public-interest text controls into ordinary campaign production. Anthropic is introducing global text watermarking for future Claude models, OpenAI is publishing training-content summaries and using layered provenance for generated media, and Microsoft is changing governance, contracts, policies, and product controls. Agencies should classify their role for each system, preserve provenance through the final export, document substantive human review, and assign editorial responsibility before publication. Europe may influence global practice less through legal imitation than through vendor features and multinational approval workflows that become easier to run as a single standard.
Japan wants providers to reveal how their models were built. Singapore is examining who owns, infringes, or invents what comes out. Between them, provenance stops being an engineering detail and becomes a business requirement.
Japan and Singapore are approaching AI intellectual property from different points in the same commercial chain. Japan’s nonbinding Principles Code asks generative AI providers to disclose how models were designed, trained, and supplied with data. It also creates conditional channels through which rights holders and users can ask whether material from an identified online location entered a model’s training data.Singapore’s consultation follows the issue through lawful access, infringing output, human authorship, inventorship, and AI-generated patent prior art. It recognizes that responsibility may shift among model developers, deployers, and users, while ownership of an AI-assisted work can depend on evidence of meaningful human contribution.Together, the initiatives push provenance into procurement, contracting, asset protection, and corporate diligence. A business seeking to commercialize an AI-assisted design or invention may need to establish where the underlying material came from, how the output was produced, and which human decisions created the protectable value. Possession of the result will no longer be enough when the chain of title cannot be reconstructed.
Moby-Dick Failed the AI Test or the Test Failed Moby-Dick
A detector can be statistically excellent and still be wrong about the page in front of it.
An AI detector can perform extremely well across a benchmark and still produce a spectacularly wrong result on an individual document. In a client project examining Pangram, Originality, and GPTZero, a block-by-block scan of Moby-Dick produced a cumulative 44% AI result. A newly written human passage about an invented subject then received a 100% AI score from Pangram, while a Claude rewrite designed to evade detection reduced that assessment to 74%.Those results expose what detector scores actually represent. Modern commercial detectors do not establish authorship or retrieve a hidden record of who wrote a passage. They classify linguistic patterns against learned boundaries between human, AI, and increasingly mixed forms of writing. Their percentages can also describe different things from one product to another.Independent research still shows Pangram performing exceptionally well, including zero false positives on a large collection of historical novels under one benchmark. That makes the Moby-Dick result an anomaly requiring controlled replication rather than evidence that the detector simply fails on old books. The larger lesson is evidentiary: detector output can be useful as a screening signal, but provenance, drafts, revision histories, and observed writing processes remain stronger evidence of authorship when the consequences of a positive result are serious.
EVA AI has found one job in artificial romance that still requires a person
EVA AI is advertising a $200-per-session role for a licensed therapist who would help people navigate emotionally significant relationships with AI companions. The proposal may offer legitimate support, since attachment to synthetic characters can produce real jealousy, anxiety and dependency. Independent research also indicates that companion chatbots can provide immediate relief from loneliness, while heavier or longer-term social-chatbot use may accompany greater emotional isolation and dependence.The unresolved issue is whether the therapist would operate independently or become an extension of the product. EVA AI’s public listing does not explain clinical privacy, referral authority, jurisdictional limits or whether recurring harms would influence product design. Its supporting survey is vendor-produced and has no publicly available methodology. A therapist could help users manage the consequences of chatbot attachment, but meaningful mitigation would also require the company to reconsider the engagement and monetization systems that cultivate it.
AI Search Is Taking Traffic—but Sending Better Customers
Brainlabs data across 54 advertisers shows a smaller acquisition stream that behaves as if it has stronger intent—leaving marketers to work out whether that engagement is worth more than the traffic it replaces.
Brainlabs data covering 54 advertiser clients over a 14-month window beginning in January 2025 shows organic sessions falling 10.5 percent while AI-platform referrals rose 163 percent. AI referrals remained far too small to replace the lost traffic, yet visitors from ChatGPT, Copilot, Gemini and Perplexity generated key events at 1.5 times the rate of organic-search visitors.The results varied by sector. Fitness, financial technology, insurance and consumer packaged goods experienced some of the largest declines, while retail, beauty and entertainment were less affected. The difference appears connected to search intent: AI summaries can satisfy many educational and comparison queries without a click, whereas transactional searches still give consumers a practical reason to visit a merchant.Higher engagement does not establish that AI referrals produce more profitable customers. Brainlabs’ key events included purchases and newsletter registrations, but also lighter actions such as reaching the bottom of a page. Marketing teams therefore need to connect referral activity with qualified leads, revenue, deal value and retention.Traffic remains economically important, but it can no longer serve as a complete judgment of search performance. Agency reporting must show both the volume lost and the value created by the visitors who remain. SEO retainers will also need to account for AI visibility, citations, and downstream business activity without replacing old traffic metrics with opaque AI scores.