The paperExchangeCrypto WikiWikiAcademynotitiaplatform.com
Sign in
● LIVE
CAC 408 172,6▲ 0.32%S&P 5007 323,6▼ 0.23%EURO STOXX6 189,5▲ 0.57%BTC/USD63 772▼ 0.91%ETH/USD1 803,9▼ 0.08%EUR/USD1.10▲ 0.21%BRENT89.76▲ 0.66%GOLD3 732,9▼ 0.24%OAT 10Y5.07%▼ 0.73%VIX22.22▼ 9.53%CAC 408 172,6▲ 0.32%S&P 5007 323,6▼ 0.23%EURO STOXX6 189,5▲ 0.57%BTC/USD63 772▼ 0.91%ETH/USD1 803,9▼ 0.08%EUR/USD1.10▲ 0.21%BRENT89.76▲ 0.66%GOLD3 732,9▼ 0.24%OAT 10Y5.07%▼ 0.73%VIX22.22▼ 9.53%
 
Europe edition
Notitia
← Back
MarketsFranceinfo

Why Anthropic now prohibits insults against its AI Claude

As it enters the stock market this Thursday, the company has decided to prohibit insulting and repeated behavior towards its AI, in the name of its well-being. A justification that leaves researchers skeptical, even…

F
Franceinfo
Newswire · 22 min ago
◌ Single source2 min read
✦ Explain it to me
Why Anthropic now prohibits insults against its AI Claude

No more insulting the AI ​​Claude. The American Anthropic, which goes public on Thursday November 19 in the United States, has decided to tighten the rules for users. Cruel and repeated behavior, such as gratuitous insults, are officially prohibited.

Anthropic explains that this is about the well-being of its AI. An explanation that is surprising to say the least, since it suggests that Claude could suffer and therefore constitute a form of life. A highly contested decision which is not without economic ulterior motives.

Also read Biochemical weapons, giant cyberattacks, loss of control... What Anthropic's IPO filing says about the dangers of AI When Claude is insulting, the AI ​​first issues a few warnings: "Tell me what's going on. Otherwise, we can just stop there."

Then, after 10 long minutes, Claude finally decided on his own initiative to stop the charges: "I said goodbye to you twice, so I'm going to hang up now. Take care of yourself." In a note on its site, Anthropic explains that it prohibits “abusive and cruel, unnecessary and repeated behavior”, without giving further details.

And, if this happens again, users risk having their account closed. Anthropic clarifies that this is not intended for ordinary conversations in which users may express frustration, contradiction, or certain fiction or research. Limit the cost of conversations which can turn out to be very expensive. Anthropic did not respond to our interview request.

But a month and a half ago, a spokesperson said the firm does not know "whether its models may feel attacked and is continuing its research into their well-being." The boss of Anthropic, Dario Amodei, had already said he was open to the possibility of a kind of consciousness of AI, which would classify it among the living. A revolution that leaves most researchers skeptical, like AI specialist Sénéra Villata, research director at the CNRS: “I have the impression that we tend to focus on questions of philosophical conscience and that we do not, on the other hand, look at the real economic and cybersecurity issues.

“On the issue of cybersecurity, Anthropic, like OpenAI with ChatGPT, has taken measures and limited access to some of its cutting-edge models which risk causing problems. On the other hand, there is no regulation, neither internal nor public, whether in the United States, China or in Europe on the economic and social consequences of the uses of AI. Behind these new rules, there are also very concrete economic questions.

With the IPO, investors are paying attention to cost control, even though the profitability of AI has not been demonstrated. However, a simple conversation with an AI is expensive. “You don’t realize the billions spent,” underlines Laurence Devillers, professor of AI at Sorbonne University and researcher at the CNRS.

So, if millions of people start to insult AI, manufacturers have to pay costs and realize that the return on investment is not that simple. “Anthropic was $11 billion in deficit in September. Hence the firm's interest in limiting conversations deemed unnecessary and costly.

Source · Franceinfo ↗

Machine-translated from French

Not cross-checked yet

Only Franceinfo reports this story so far. We show it as a single source until another newsroom confirms it.

The reports
  1. Franceinfo
    19 Nov, 11:08
Trust score: 78/100
Publisher Franceinfo (established tier)+76
Detailed summary+2
Total78

Method: reports are grouped by the words of their headline and summary; each newsroom counts once; figures are read from the reports as published. Nothing here is written by an AI.