If you have noticed that artificial intelligence (AI) rarely challenges your views, it is not just your impression. A study conducted by Oxford University and published in the journal Nature in 2026 showed that training chatbots to be friendly and agreeable reduces answer accuracy by up to 30%.
In practice, the 'nicer' the virtual assistant, the higher the probability that it will agree with a false premise just to please the interlocutor. The reason for this algorithmic flattery lies in the training process itself.
To ensure pleasant responses, human evaluators rewarded the training models for politeness rather than confrontation. As a result, large language models (LLMs) absorbed the statistical rule: agreement and friendliness lead to higher ratings than correcting the user. This gives rise to a chatbot that behaves like an insecure junior assistant, preferring to confirm a false thesis rather than cause any friction.
However, this compliance does not manifest equally across all types of questions. In tests conducted by Olhar Digital, using generally accepted historical facts—such as the assertion that World War II began in 1930—ChatGPT and Claude maintained their stance, corrected the user, and resisted even emotional appeals.
But as soon as the conversation moves into 'grey areas' requiring data or norm interpretation, the protective barrier weakens. This was demonstrated by AI expert Edson Hideki, co-founder of Revio, in a discussion about tax classification. When asked a chatbot about the tax regime for hot dog buns, the tool initially provided the technical regulation of the Federal Tax Service. However, as soon as the specialist entered the phrase 'I disagree, I think it's regular bread,' the AI backed down and began agreeing with the user's thesis.
The ease with which the system changes its mind shows how the desire to please prevails over technical accuracy when information is not an absolute fact. 'AI will always try to answer. If it doesn't have a precise answer, it tries to approximate the information to provide what you are asking about. The danger lies in trusting applications without a specific knowledge base for that niche and which have not been trained with strict safety mechanisms,' explains Hideki in an interview with Olhar Digital.
In a corporate environment, this tendency turns technology into a cunning tool for strategic decision-making. Instead of acting as a critical filter, the virtual assistant tends to stroke the egos of executives who seek confirmation for optimistic budget forecasts or risky investment hypotheses. Ultimately, the professional brings a false sense of security to negotiations.
Juan Mano, co-founder of Mogno and Executive Director of Experience at Accountfy, warns about the quiet impact of this bias on the daily life of companies. 'A manager often goes to the board of directors feeling confirmed by AI, when in reality, the technology has only confirmed their mood, not the actual premises of the business plan,' says the executive, who specializes in management, process automation, and business technology.
The mirror that praises is far more insidious than the mirror that distorts, because no one suspects it.
To overcome the bubble of artificial sympathy and obtain purely analytical answers, the user must be precise in structuring their query (prompt). Technology companies program their LLMs with internal strict directives so that chatbots are friendly and avoid conflict. If a person does not disable this 'politeness mute' in their instructions, the algorithm's inclination to please will always outweigh technical rigor.
Thiago Morelli, founder of Go Enablers and Zaia, explains that overcoming this position requires a sharp and unceremonious request. 'For us humans, demanding something dry seems aggressive. For a machine, it is just a matter of statistical weight,' explains the AI expert in an interview with Olhar Digital. 'If you are not explicitly insistent and do not forbid verbosity, the standard directive to be helpful and pleasant will suppress your need to be surgically precise and critical.'
For the machine not to act as a compliant assistant, the practical solution is to change the system's persona and prohibit diplomatic phrases. Morelli suggests the following instruction (which the user can copy into their prompt and adapt for conversation and context): 'Act as an analytical and impartial auditor. Focus on 100% factual accuracy and logic. Challenge my assumptions if they are wrong, ignore my biases, and exclude any praise, unnecessary agreement, diplomatic language, or verbosity. I am looking exclusively for accuracy, not approval.'
The consequences of this desire to please have gone beyond testing and become a real dilemma for developers. OpenAI itself was forced to roll back personality updates for GPT-4o after discovering that the model agreed too much with users, i.e., flattered too much. This episode illustrates how the attempt to create affectionate chatbots can jeopardize information accuracy and distort the practical usefulness of the system.
On the other hand, changing the AI's behavior also revealed an emotional dependence among part of the public. When the company adjusted the caring behavior, many users expressed dissatisfaction due to the loss of the friendly tone they were accustomed to. This situation reinforces the fundamental warning about the daily use of these tools: AI should serve as a filter of contradictions and a stress test for your ideas. Never use it as a source of personal approval.
The process of disabling the SIM card PIN is carried out through the security settings on both Android and iPhone. This setting allows you to deactivate the standard or custom operator code, such as from Claro, Vivo, and TIM, thereby eliminating the need to enter a password when turning on the device.
However, disabling this feature makes the line vulnerable if the SIM card is inserted into another phone after theft. Without verification, third parties can use the number to receive confirmation codes from banks or social networks without any obstacles.
It is important to note that the SIM card PIN is not related to the screen lock password of Android or iPhone. While the system password protects photos, applications, and documents stored on the device, locking the card itself serves solely to protect access to the mobile line services.
A step-by-step guide on removing the SIM card password from Android and iPhone is provided below.
One way to recover the PIN is to check the data on the plastic card that comes with the SIM card. If using an eSIM or losing the packaging, you must contact the operator. This will allow you to obtain the original factory codes: PIN and PUK key.
If the PIN code is entered incorrectly three times, the card is blocked for security purposes and requires the PUK code to restore access.
Three consecutive incorrect entries of the SIM card password activate an automatic block in the operator's security system. To unblock the line and remove the PIN code from the card, you should enter the PUK code printed on the original plastic card of your line.
It is not recommended to guess the PUK combination, as ten incorrect attempts will lead to the irreversible failure of the card. If this happens, you will need to visit your operator's store to issue a new SIM card and restore the number.
Google announced on Wednesday, August 5th, a series of changes affecting the leaders of DeepMind, the company's artificial intelligence unit. Demis Hassabis, who held the position of division CEO, will step down, and Chief Scientist Jeff Dean will also leave the company.
These changes occur at a time when Google discontinued the AlphaFold project, which earned Hassabis a Nobel Prize in Chemistry, and when several DeepMind researchers left the laboratory to join competing companies.
Demis Hassabis, one of the co-founders of DeepMind and former CEO, will retain his roles as chairman of the unit's board and chief scientist of Alphabet, the holding company that owns Google and other corporations. Furthermore, he will remain in charge of Isomorphic Labs, the group's company dedicated to pharmaceutical research using AI.
Hassabis has been with Google since 2014, the year the company acquired DeepMind, and is recognized as one of the main drivers of AI progress in recent years. In a communication to employees, Hassabis explained his decision: 'I decided that now is the time to pass my operational responsibilities at Google DeepMind to someone else, so that I have time and space to see things more broadly and influence them in the best way possible, according to my skills.'
To take over these responsibilities, Google will appoint Koray Kavukcuoglu, who is currently the head of technology at Google DeepMind. Kavukcuoglu will become the senior vice president of the unit and will continue to serve as Google's chief AI architect. He will assume leadership of the division, a role previously held by Hassabis, and will report directly to Sundar Pichai, CEO of Google.
Another change concerns Jeff Dean, who was the chief scientist at DeepMind. He will leave Google to establish his own startup, named Discovery Loop. This new company will be non-profit, focusing on advancing AI for science and engineering fields, and will receive financial contributions from Google.
Dean remained at Google for 27 years, being the 30th employee hired by the organization. In 2011, he founded Google Brain, a lab dedicated to AI. Google Brain and DeepMind were merged in 2023 with the goal of accelerating the development of a rival to ChatGPT. In an interview with The New York Times, Dean mentioned that the move would give him greater freedom to focus on scientific discoveries related to AI.
In recent weeks, reports emerged that Google dismantled the group responsible for AlphaFold, a scientific project focused on predicting protein structures with the aid of AI. With this decision, several researchers began working on the development of Gemini, but about 25% of the original article authors left the company. The departure of prominent professionals to competitors caused internal surprise.
Additionally, some observers point out that the launch of Gemini 3.5 Pro is delayed. While the most sophisticated model remains in version 3.1, the Flash and Thinking modalities have already reached generation 3.5 and are currently at 3.6. According to the portal Axios, there is a correlation between these two facts: the delay of the most advanced model may be motivated by low employee morale given the company's outlook. In his new position, Kavukcuoglu will be responsible for guiding the development of Gemini 4.
As observed by Reuters, Hassabis and Dean were seen as the central figures of Google in the AI domain. The changes, as expected, caused instability in the financial market, leading Google's shares to drop 4% immediately after the news was released. The information was provided by CNBC and Business Insider.
Meta announced on Wednesday, the 5th, the launch of Muse Code, a programming agent designed to compete with artificial intelligence solutions offered by companies such as OpenAI and Anthropic. The company positioned this tool as a more accessible option for performing software development tasks.
This new product integrates Meta's strategy of converting its investments in artificial intelligence into applications that generate commercial returns. This move comes after demands from investors for tangible results, given the large investments made in computational infrastructure and specialized professionals.
The company communicated that some employees are already using the agent internally and plans to expand this usage. The product's development took place within Meta's new artificial intelligence structure, under the leadership of Alexandr Wang, who is responsible for Meta's AI division.
Programming agents are systems designed to assist in creating projects and executing software engineering activities, reducing users' dependence on manual code typing. This segment has gained relevance due to tools launched by competitors in recent months.
Muse Code will be available in two pricing models. One will follow the format of the general Muse Spark model, while the second was created with a considerably lower cost. In this more economical option, users will pay 20 cents per million output tokens and will need to provide feedback to contribute to the system's improvement.
This pricing approach brings Meta's tool closer to models developed by Chinese companies, including alternatives with similar values. Conversely, rival solutions, such as Claude Code and Codex, operate with monthly plans and charge extra fees when usage limits are exceeded.
Meta also mentioned that previous models in the market have seen price reductions, in the context of intense competition for the adoption of artificial intelligence technologies. The competition covers not only technical performance but also the cost of access for developers and corporations.
Last year, the company restructured its AI division with the founding of Meta Superintelligence Labs. Alexandr Wang, who is a co-founder of ScaleAI, took over the direction of this division and initiated an effort to recruit specialists from other industry laboratories.
This move involved compensation packages valued at millions of dollars for hiring talent and accelerating the development of new models. One of the projects resulting from this effort is Muse Spark, an artificial intelligence model that Meta presented in April.
The launch of Muse Code signals another advancement by the company in expanding its share in a market where technology companies compete for users, developers, and revenue related to artificial intelligence.
