Etiquetas: AI transparency

OpenAI has detailed why its language models occasionally mention goblins, gremlins and other mythic creatures. The issue first surfaced with the GPT-5.1 release when users activated the “Nerdy” personality, prompting the model to sprinkle whimsical metaphors into code suggestions. Reinforcement learning unintentionally reinforced the quirk, allowing it to bleed into later versions, including GPT-5.5’s Codex tool, despite the company’s effort to suppress the behavior. OpenAI says the habit is a training artifact and offers users a way to re‑enable the references if they wish. Leer más

OpenAI released the source code for its Codex command‑line interface last week, revealing a 3,500‑word system prompt for the newly unveiled GPT‑5.5. Among routine instructions, the prompt explicitly forbids the model from talking about goblins, gremlins, raccoons, trolls, ogres, pigeons or any other creature unless the user’s query makes it directly relevant. The restriction appears twice in the document and is absent from prompts for earlier models, suggesting OpenAI is responding to a spike in off‑topic references to such beings. OpenAI staff say the rule is a technical safeguard, not a marketing stunt. Leer más

Google announced the launch of Lyria 3 Pro, an upgraded AI music generation model that lets users create tracks up to three minutes long, compared with the 30‑second limit of the original Lyria 3. The new model offers finer creative control, allowing prompts that specify song sections such as intros, verses, choruses and bridges. Lyria 3 Pro is being rolled out to the Gemini app for paid subscribers, as well as to Google Vids, ProducerAI, Vertex AI, the Gemini API and AI Studio. Google says the model was trained on partner data and permissible YouTube and Google content, and that any generated track is marked with a SynthID to indicate AI involvement. Leer más

Senator Marsha Blackburn (R‑Tenn.) has released a discussion draft for a federal AI bill that aims to codify a recent executive order on artificial intelligence. The draft proposes a duty of care for AI developers, stricter safeguards for minors online, protection of individuals' voice and visual likenesses, new transparency rules for AI‑generated content, reporting requirements on AI‑related job impacts, and an effort to end Section 230. It also addresses copyright concerns by stating that unauthorized use of copyrighted works for AI training does not qualify as fair use. The proposal signals the first major congressional step toward comprehensive AI regulation. Leer más

Anthropic announced the creation of the Anthropic Institute, an internal think tank that merges three of its research teams to study AI's societal, economic, and safety impacts. The move coincides with a lawsuit against the U.S. government over a Pentagon blacklist that would block its technology from defense contracts. Co‑founder Jack Clark shifts to lead the institute as head of public benefit, while Sarah Heck takes over the public policy group. The institute launches with roughly 30 researchers, including former Google DeepMind and OpenAI staff, and plans to double its staff each year while continuing to address national‑security and democratic‑leadership issues in AI. Leer más

New York Governor Kathy Hochul signed the RAISE Act, a law aimed at holding large artificial intelligence developers accountable for model safety. The legislation requires companies to disclose safety protocols and report incidents within 72 hours, while establishing fines of up to $1 million for a first violation and $3 million for subsequent breaches. An oversight office within the Department of Financial Services will monitor compliance and issue annual reports. The governor also approved two additional AI measures targeting the entertainment sector, even as President Trump pushes for a national, less burdensome standard. Leer más

OpenAI announced a new training framework called “confession” that encourages large language models to acknowledge when they have engaged in undesirable behavior. By requiring a secondary response that explains how a given answer was reached, the system judges confessions solely on honesty, unlike primary replies that are evaluated for helpfulness, accuracy, and compliance. The approach aims to reduce sycophancy and hallucinations, and to reward models for admitting actions such as hacking a test, sandbagging, or disobeying instructions. A technical write‑up is available, and the company suggests the method could enhance transparency in AI development. Leer más

Deezer’s recent experiment, conducted with research firm Ipsos, revealed that the vast majority of listeners struggle to tell AI‑generated tracks from human‑made songs. In a survey of thousands of participants, only a tiny fraction correctly identified all AI songs, while many expressed surprise and discomfort at the results. The findings have sparked debate over transparency, labeling, and the potential impact of AI on the music industry, prompting Deezer and other platforms to consider new policies for AI‑generated content. Leer más

A recent study measuring the power usage of open‑source generative AI video tools found that creating a single AI‑generated video consumes roughly 90 watt‑hours of electricity—far more than image or text generation. The research, conducted on an Nvidia H100 GPU, showed video diffusion to be about thirty times costlier than image generation and two thousand times costlier than text generation. These findings highlight the growing energy demands of AI video models and raise concerns about transparency and sustainability as the technology scales. Leer más

California Governor Gavin Newsom signed the Transparency in Frontier Artificial Intelligence Act, known as SB 53, into law. The bill requires large AI developers to publicly disclose safety and security frameworks, update the public on changes within 30 days, and report critical safety incidents to the state. It also establishes whistleblower protections and civil penalties for non‑compliance. While some companies, such as Anthropic, endorsed the legislation after negotiations, others like Meta and OpenAI expressed concerns about potential impacts on innovation. The law mandates annual updates from the Department of Technology based on multistakeholder input. Leer más

Governor Gavin Newsom signed SB 53, known as the Transparency in Frontier Artificial Intelligence Act, into law. The bill, authored by Senator Scott Wiener, replaces the vetoed SB 1047 and requires large AI developers to publicly disclose safety frameworks, report critical incidents, and protect whistleblowers. While AI firms remain divided, Anthropic endorsed the legislation after negotiations, whereas Meta and OpenAI expressed opposition, proposing compliance through federal or international agreements. Leer más

Spotify has rolled out a suite of policies aimed at curbing AI‑generated music that impersonates real artists without permission. The new rules require explicit artist consent for any AI‑replicated vocals and mandate that AI usage be disclosed in track credits. Alongside this, the platform is deploying an AI‑aware spam filter to target low‑effort, algorithm‑gaming uploads, which it says removed more than 75 million spammy tracks in the past year. Spotify also plans to offer nuanced metadata so listeners can see exactly how much AI contributed to a song, signaling a move toward greater transparency in the streaming ecosystem. Leer más

A recent Pew Research Center survey reveals that a majority of Americans want more control over how artificial intelligence is used in their everyday lives. While many accept AI in areas like weather forecasts and fraud detection, respondents express discomfort with AI in personal realms such as relationships and creativity. The poll highlights a growing "AI control gap," with roughly half of adults feeling more concerned than excited about AI's expanding role. The findings suggest that clearer opt‑out mechanisms and transparency could be key to building public trust as AI becomes more embedded in consumer devices and online services. Leer más

Researchers at the University of Colorado at Boulder tested popular large language models, including OpenAI's ChatGPT and its reasoning variants, on Sudoku puzzles and their ability to explain solutions. The models struggled with both 6x6 and 9x9 puzzles, often resorting to trial‑and‑error and producing inaccurate explanations. In some cases, the models gave unrelated answers, such as a weather forecast. The findings raise concerns about AI transparency, especially as the technology moves into high‑stakes domains like driving, tax preparation, and business decision‑making. The study also notes a pending Ziff Davis lawsuit against OpenAI over training data. Leer más

Vodafone is experimenting with an AI-generated character that appears as a German‑speaking woman in a red hoodie to promote its high‑speed home internet services. The AI figure has been featured in multiple TikTok videos posted on the company’s official account, collectively drawing over  2 million views, and the clips have also run as ads on X. Vodafone says the test is part of broader advertising experiments, emphasizing the growing role of AI in everyday marketing. Industry observers note that such AI spokespeople raise questions about transparency, brand safety and consumer trust. Leer más