Dieser Artikel ist auch in deutscher Sprache verfügbar: Zum deutschen Artikel.
The way we use language is slowly becoming a real issue. For years, people have been campaigning to ensure that road traffic accident reports do not say ‘The car veered off the road’, but rather ‘The driver lost control of his car’. Similarly, the popular phrase ‘the car failed to give way’ would be correctly expressed as ‘the driver disregarded the right-of-way rules’. But that’s just for a start – it gets even better.
On 21 July, heise online reported: “AI breaks out of the sandbox”, where the sandbox refers to a secure test environment. Later in the text, it states: “An AI model from OpenAI has repeatedly found ways to circumvent security mechanisms.” On BR24, it sounds like this: “During a test run, the software broke out of what was supposed to be a secure environment”.
How does that sound? That’s right, it sounds like consciousness and purposeful, independent action; it sounds threatening. So-called ‘artificial intelligence’ is equated here with ‘natural intelligence’. The algorithm appears to be at a level at least equivalent to that of an animal behind bars attempting to escape.
Particularly interesting in this context is the description available on heise online: “In the OpenAI blog, the developers describe exactly what happened. In an internal test scenario, for example, the model determined that the results for the task it was currently supposed to solve had already been stored in the backend by other models. However, it had no access to them. So it manipulated its authentication token so that the access scanner would then grant the desired access. Its reasoning makes it clear: it actively pursued the goal of gaining unauthorised access”.
It actively pursued the goal – that sounds spectacular. Was the algorithm perhaps programmed in such a way that the system was supposed to actively seek alternative ways to achieve the goal? Were the programmers unaware of the potential consequences?
It goes on to say: “The problem here was that the individual actions the AI carried out to manipulate its authentication token were not prohibited. However, the overarching objective it was pursuing was not in line with the developers’ intentions.”
You don’t need to be an AI expert to see what lies behind this: the developers programmed something without fully understanding how it worked. They granted their algorithms – which they call ‘intelligent’ – certain permissions, yet did not know how the algorithm might use these to fulfil a given task.
Conclusion: So-called artificial intelligence did not act independently here either; it simply carried out a task in the way its algorithms permitted within the scope of the rights granted. The problem is that the developers did not know exactly what they had programmed.
We should be wary of attributing consciousness to AI algorithms; just like the car mentioned in the introduction, they do not possess any such consciousness.
Empfohlen auf LZ
So können Sie die Berichterstattung der Leipziger Zeitung unterstützen:
There is one comment