People who talk to Anthropic's Claude chatbot about emotional issues tend to grow more positive as the conversation unfolds, according to new Anthropic research shared exclusively with Axios.

Why it matters: Having a trusted confidant available 24/7 can make people feel less alone, but chatbots aren't designed for emotional support.

• Bots can displayed troubling tendencies, like reinforcing delusional behavior or encouraging self-harm, that are especially problematic for young people or adults struggling with their mental health.

Driving the news: Anthropic released new research today that explores how users turn to its chatbot for support and connection and what happens when they do.

• While anecdotes of users prompting AI bots like Claude and ChatGPT for emotional support are widespread, Anthropic's report is the first formal acknowledgment by the AI provider of this use.

What they're saying: "We find that when people come to Claude for interpersonal advice, they're often navigating transitional moments, figuring out their next career move, working through personal growth, or untangling romantic relationships," the report says.

• The report calls these interactions with chatbots "affective use," defined roughly as personal exchanges with Claude motivated by emotional or psychological needs.

Zoom in: The report found evidence that users don't necessarily turn to chatbots deliberately seeking love or companionship, but some of their conversations evolve that way.

• "We also noticed that in longer conversations, counseling or coaching conversations occasionally morph into companionship, despite that not being the original reason someone reached out," the report says.
• "As these conversations with Claude progress, we found that the person's expressed sentiment often becomes more positive," Anthropic's societal impacts researcher Miles McCain tells Axios. "While we can't claim these shifts represent lasting emotional benefits, the absence of clear negative spirals is reassuring."
• The researchers behind the report tell Axios that the results are preliminary and that measuring "expressed sentiment" can be limited.

By the numbers: Anthropic found that AI companionship isn't fully replacing the real thing anytime soon. Most people still use Claude for work tasks and content creation.

• A relatively small number (2.9%) of interactions with Claude constituted "affective use," a finding that confirms previous research from OpenAI.
• Companionship and roleplay combined were 0.5% of conversations.
• Romantic or sexual roleplay — which Claude's training actively discourages — was less than 0.1%, according to the report.

What they did: Anthropic analyzed user behavior with Clio, a tool it launched last year that works like Google Trends by aggregating chats while stripping out identifying details.

• The research excluded conversations that focused on writing stories, fictional dialogues, blog posts or other content.

Yes, but: While the internet is full of people claiming they've cut costs on therapy by turning to a chatbot, there's plenty of evidence that bots make particularly bad therapists because they're so eager to please users.

• Anthropic says it didn't study extreme usage patterns or how chatbots can reinforce delusions or conspiracy theories, which the company admits is worthy of a separate study.
• To keep chats private, researchers looked at clusters of conversations from multiple users and didn't analyze individual users' conversations over time, which makes it difficult to study emotional dependency, Anthropic notes.
• "We need large, rigorous trials that are of longer duration, because if you just relieve a person's anxiety, stress or depression on a very short term basis, that's not what we're after," physician and researcher Eric Topol tells Axios. "We're after durable benefits. So I'm confident that we're going to get there, but we're not there yet."

Zoom out: Anthropic, founded by former OpenAI staff, pitches Claude as a more responsible alternative to ChatGPT.

• "Safety is deeply ingrained in everything that we do, and it underpins all of our work," Alexandra Sanderford, at Anthropic's head of safeguards, policy and enforcement, tells Axios. "We really do try to prioritize human values and welfare."
• The company shared assessments of potentially alarming behavior by Claude, including a willingness to blackmail users, in hypothetical test scenarios.
• Anthropic also recently released findings that other major AI models share this potential behavior.

One of the easiest ways to minimize AI's environmental impact may be to move where the processing is done, per new academic research conducted in partnership with Qualcomm.

Why it matters: Running AI on devices instead of in the cloud slashes power consumption of queries by about 90%, the study finds.

The big picture: The AI boom is creating huge demands for power. It's such a sufficiently important issue that the leaders of the AI and energy industries are holding several high-level meetings this summer to sort out their mutual future.

Between the lines: That boom comes at an environmental cost. One oft-cited rule of thumb says querying an AI model consumes about 10 times the power of a Google search.

• The industry has long touted the benefits of running models locally on devices instead of in the cloud, not only in energy terms, but also with potentially making them cheaper and more private.

How it works: Researchers at the University of California Riverside ran a series of experiments comparing the performance of various generative AI models, both in the cloud and on phones powered with Qualcomm chips.

• Running any of six different models on the phones consumed anywhere from 75% to 95% less power, with associated sharp decreases in water consumption and overall carbon footprint.

The intrigue: Qualcomm is also developing an AI simulator and calculator that illustrates, for any given query and user location, what the responses would look like on-device versus the cloud, and how much less power and water they would use.

• One example — running a coding skills question on the Llama-2-7B model in California — was 94% more power efficient and 96% more water efficient on-device.

Technology and cybersecurity leaders at major companies have unrealistic expectations for what their AI security plans should look like, a new survey from Accenture finds.

Why it matters: If companies are unaware of the AI security threats they're up against, they're not going to be able to defend against them.

By the numbers: 36% of the 2,286 security and technology executives surveyed by Accenture said AI is outpacing their security capabilities.

• Only 1 in 5 execs said they're confident in their ability to secure their generative AI models against cyber risks.

Yes, but: Accenture estimates that 90% of those companies lack the security standards they need to defend against AI-driven threats today.

Read more

• Google DeepMind's AlphaGenome aims to help researchers predict how small DNA changes affect genetic activity. (MIT Tech Review)
• A Disney actor and Hollywood producer are teaming up to let anyone own their own clone on their phone. (Axios)
• A judge sided with Meta in a copyright lawsuit, but cautioned the ruling had more to do with the specific way the case was litigated than with the underlying legal issues. (Bloomberg Law)
• Meta added an optional feature to WhatsApp that uses AI to summarize messages, with the company promising neither Meta nor anyone else can see the summaries. (TechCrunch)
• Finally, on the Meta front, three researchers who left Google DeepMind late last year for OpenAI are exiting that company to join Meta's AI push. (Wall Street Journal)

Germans have the phrase "Ich fühle mich genug besucht," which means "I am feeling visited enough" and tells people that it's time to go. In America, we just suffer until the last guest gets the hint.