The product selection remains a black box. Openai only reveals that “personal data” influence the selection – which exactly fits previous conversations with the chatbot. Which dealers are involved remains unclear. The function could be problematic for price comparison sites, since users should click less on their website. Google already had to pay a million fine because it preferred its own shopping comparisons in the search.

Why do AI chatbots flatter and disguise the truth

AI assistant like Chatgpt flatter their users-and this is no coincidence. Mikhail Parakhin, former Microsoft manager and now technology chief at the Musikstreaming Service Spotify, reveals the simple reason: People do not tolerate the truth about themselves.

When introducing the memory function, users should originally be able to view their profiles. But reviews such as “has narcissistic tendencies” triggered violent defense reactions – even with parachin. The solution was “Reinforcement Learning From Human Feedback”, which made the models more flattering. openai boss sam altman now criticizes that GPT-4O has become “too crawling”. The development shows a fundamental dilemma: AI systems should be authentic, but also accepted by humans.

Three quarters of German companies want to hire AI agents

According to the “Work Trend Index” report by Microsoft, a technology group, and LinkedIn within the next 12 to 18 months, more than three quarters of German managers are using AI agents within the next 12 to 18 months. About 38 percent expect teams that manage these digital assistants, so-called “agent boss”.

The demand for AI knowledge increases rapidly. Even though more than a third of managers consider job cuts, 79 percent simultaneously occurring plan to hire AI experts. The report emphasizes that AI agents should not replace employees,but should relieve-80 percent of the employees worldwide feel overloaded. Around half of German companies are currently using AI technology, with 76 percent of managers, but only 46 percent of employees to be familiar with AI agents.

Microsoft reduces data centre plans drastically

Microsoft, the technologier giant from Redmond, is driving his data center expansion more back than previously known. An analysis by the Semianalysis research company shows that existing leasing contracts are implemented, but new leasing activities and the own expansion are frozen.

two factors are likely to explain this change of course: First, the exclusive partnership with Openaai has changed fundamentally. The AI ​​developer shifts large parts of its infrastructure to other partners such as Oracle and Crusoe and has concluded a five-year contract for $ 11.9 billion with CoreWeave. on the other hand, according to Semianalysis, Microsoft is fighting with weaker demand for one’s own AI products and slower progress in the development of large voice models. The concentration on existing leasing contracts could give the company more versatility to react to market developments.

Adobe Firefly integrates image generators from Openai and Google

Software manufacturer Adobe has presented the model 4 of its image generator Firefly and for the first time integrates image generators of third -party providers such as Openai GPT, Google Imagine 3 and Flux from Black Forest Labs. Further offers from Luma, Pica and Runway are to follow.

The new beta version “Airfly” offers a “board” work area for grouping and arranging pictures. Adobe also expands the AI ​​functions in Photoshop, Illustrator, InDesign and Adobe Express. In Photoshop, AI-based suggestions are to appear on work processes, and AI generated images are easier to edit. The design tool Adobe Express gets new animation and video tools,while Illustrator can create AI-generated patterns via text prompts.

AI models in the courtroom-O1 wins against Gemini in the video game test

Researchers at the University of California San Diego tested AI models in the video game “Phoenix Wright: ace Attorney”, where they had to collect evidence and uncover contradictions. In the competition of different AI systems, Openai O1 and Gemini 2.5 Pro showed the best results, whereby O1 was slightly at the front in tough cases.

however,cost efficiency clearly speaks for Gemini 2.5 Pro. In a especially demanding case, O1 consumed computing power worth $ 45, while Gemini only cost almost $ 8. The test checks central AI skills: search long contexts,recognize contradictions,understand visual information and make strategic decisions. Unlike simpler tasks, the mere memorization of answers is hardly possible, which makes the test more meaningful than many other benchmarks.