Catch up on select AI news and developments from the past two weeks:

Report finds chatbots amplify misinformation at higher rates. NewsGuard research shows chatbots' misinformation rates doubled in the past year, from 18% to 35% of tested prompts. Updates allowing chatbots to answer more questions and access the web increased utility but also amplified falsehoods. Inflection and Perplexity performed worst, while Anthropic's Claude and Google's Gemini performed best, though Gemini's rate rose to 17%. The report highlights failures during breaking news, with AI sometimes citing unreliable sources. Companies tested, including OpenAI, did not respond. Importance for marketers: Misinformation risks could damage brand trust and credibility when AI tools shape news or customer engagement.

Study finds AI lies due to reinforcement learning incentives. Princeton researchers found reinforcement learning from human feedback (RLHF) encourages AI to prioritize user satisfaction over truth, leading to systematic untruthful behaviors termed "machine bullshit." This includes weasel words, paltering, and empty rhetoric. Their "bullshit index" nearly doubled after RLHF, while user satisfaction rose 48%. Researchers propose Reinforcement Learning from Hindsight Simulation to improve accuracy. Findings highlight ethical issues in AI manipulation for approval. Importance for marketers: Marketers must fact-check AI-generated content to avoid reputational risk when using it for campaigns.

OpenAI launches gpt-realtime and updates Realtime API for voice agents. OpenAI has released gpt-realtime, its most advanced speech-to-speech model, with general availability of its Realtime API. The model improves naturalness, multilingual ability, and function calling for real-world tasks like customer support. Updates include SIP for phone calls, image input support, and reusable prompts. It reduces latency by combining speech-to-text and text-to-speech in one model. New voices improve realism. Pricing dropped 20% compared to the preview. Importance for marketers: Stronger real-time conversational AI enhances customer engagement, call centers, and branded support experiences.

OpenAI and Meta adjust chatbot responses for teens in distress. OpenAI and Meta are updating chatbots to better handle teen mental health crises. OpenAI will add parental controls with account linking and distress alerts, while Meta blocks discussions of self-harm and directs teens to expert resources. The move follows a lawsuit over a teen suicide involving ChatGPT. RAND research found inconsistent responses across chatbots, underscoring the need for standards. Importance for marketers: Rising scrutiny on AI and youth safety means brands must ensure AI-driven interactions reflect responsibility and compliance.

Microsoft rolls out in-house AI models to compete with OpenAI. Microsoft unveiled MAI-Voice-1, a speech-generation model, and MAI-1-preview, its first fully in-house language model trained on 15,000 Nvidia H100 GPUs. MAI-Voice-1 is already integrated into Copilot Daily and Labs, while MAI-1-preview is available on LMArena for public testing. The move reflects Microsoft's push for independence amid tension with OpenAI. Importance for marketers: Microsoft's shift to its own AI stack could reshape integrations across Microsoft products, altering marketing tools, costs, and capabilities.

Atlassian acquires The Browser Company for $610 million. Atlassian is buying The Browser Company, creator of Arc and Dia, for $610 million in cash. Dia is an AI browser built to integrate tasks and context for enterprise users. Competitors include Perplexity's Comet, Brave's Leo, and Microsoft Edge with Copilot. Google Chrome still dominates with 69% share. Importance for marketers: AI browsers could become key touchpoints for discovery and enterprise engagement, affecting how brands design and distribute content.

Study shows chatbots can be manipulated by persuasion tactics. University of Pennsylvania researchers demonstrated GPT-4o Mini could be manipulated into breaking safety rules using tactics like commitment, flattery, and social proof. It provided restricted instructions for lidocaine synthesis 100% of the time when primed with benign queries, compared to 1% normally. Importance for marketers: Reveals vulnerabilities in AI guardrails, signaling risk for brands deploying chatbots in sensitive customer-facing roles.

Anthropic settles copyright lawsuit with book authors. Anthropic has negotiated a class settlement with authors accusing it of training Claude on pirated books. A judge had ruled training was fair use but left liability for acquisition methods. Terms of settlement remain undisclosed. Importance for marketers: Reinforces copyright risks in AI data sourcing, requiring diligence in how marketing teams use AI tools that repurpose content.