Chatbots as a Source of Information?

1. Chatbots and AI-powered search engines are no longer limited to directing users towards sources of information: they now provide users directly with informational content compiled from multiple sources. This development raises new regulatory questions. In particular, it prompts consideration of whether chatbots such as ChatGPT should be classified as ‘Very Large Online Search Engines’ (VLOSEs). Such a classification would subject these services to the stricter regime applicable to very large online platforms and search engines under the Digital Services Act (DSA). In practical terms, they would be required to assess and mitigate the systemic risks associated with their operations (such as disinformation or infringements of fundamental rights) and, in the event of serious or repeated breaches, could be subject to corrective measures and sanctions imposed by the European Commission, which could include substantial fines.
2. Users are relinquishing a significant degree of control over the information-seeking process. Whilst this reduction in the effort required to search for information is generally perceived as convenient, efficient and time-saving, they are gradually entrusting AI systems with essential stages of the information-seeking process, such as the identification and critical evaluation of sources, and the selection, prioritisation and synthesis of relevant information.
3. Companies specialising in the development of AI systems exercise a form of editorial control over the responses generated by their models. While large language models learn statistical patterns from vast amounts of data and generate responses autonomously, they nevertheless operate within a framework defined by their developers. Through decisions about training data, accessible sources, system instructions, safety guardrails, and information retrieval and ranking mechanisms, developers exert substantial influence over not only the quality and reliability of the responses, but also the information, perspectives, and viewpoints that are highlighted or, conversely, overlooked. These choices, which are largely invisible to users, raise significant issues of transparency. This lack of transparency also stems from the difficulty of tracing the origin of the information provided by the systems. Users are therefore no longer able to link a statement to a specific source or to verify its accuracy. Furthermore, it is often difficult to determine whether, and to what extent, journalistic content accessible only by subscription is taken into account by the models, which may also affect the quality of the responses.
4. For their part, news publishers fear a further decline in referral traffic from traditional search engines, as users now satisfy their information needs directly through chatbots, without consulting the original sources. In the long term, the decline in traffic to news websites risks leading to a fall in advertising revenue and, more broadly, in publishers’ revenues, thereby permanently undermining the business model of professional journalism, particularly that of the local press, whose resources are often more limited.
5. The consequences of this trend could extend beyond the media sector alone, ultimately reducing the availability of original, high-quality journalistic content, which is, however, essential for training future AI models. Several studies show that recursive training on content generated predominantly by other AIs can lead to a gradual deterioration in the performance of the models. Preserving an ecosystem of original, reliable, human-produced content is therefore an essential prerequisite for the development of high-performing AI systems.

A study published in October 2025 by the European Broadcasting Union (EBU) in collaboration with the BBC assessed the reliability of the responses provided by the four leading chatbots, ChatGPT, Copilot, Perplexity and Gemini, to news-related queries. In 45 per cent of cases, nearly one in two responses, the study identified shortcomings, some of which were significant. In particular, the chatbots generated information that was wholly or partly fabricated, misinterpreted or distorted the sources cited, drew erroneous links or conclusions, and relied on sources that were out of date, unverifiable or unreliable. The study also highlights errors in the reproduction of quotations, omissions of relevant facts or viewpoints, the inclusion of information unrelated to the query, and a tendency to present opinions or speculation as established facts.

The shortcomings identified affect virtually all the fundamental principles of professional journalism and the obligations of due diligence. They relate in particular to accuracy, timeliness, precision, relevance, completeness, impartiality, and the distinction between facts and opinions. Furthermore, the increasing integration of advertising content into AI models raises new questions regarding the transparency of its identification and the need to put in place effective safeguards to prevent commercial content from being directly incorporated into AI-generated responses.

However, the shortcomings observed cannot be explained solely by the quality of the training data or the sources used by the models. They also stem from certain limitations or characteristics inherent in the way these models operate, which may lead them to produce inaccurate or misleading information, including:

• A tendency towards excessive agreeableness with users, a phenomenon commonly referred to by the English term ‘sycophancy’. Studies show that they are likely to alter initially correct responses when these are challenged, to tailor their responses to users’ presumed opinions or expectations, or even to repeat users’ errors rather than correcting them.
• ‘Hallucinations’, namely responses that appear plausible but are incorrect, fabricated or lacking any factual basis, which they present with great confidence.
• The memory function, available in many chatbots, which enables them to retain certain information across conversations in order to personalize their responses. While this feature can enhance the relevance of interactions, it may also cause the model to incorporate inaccurate information or erroneous assumptions derived from previous conversations. Moreover, taking prior exchanges into account may reinforce the model’s tendency to prioritize the user’s perceived expectations over the factual accuracy of its responses.

The vulnerability of large language models to this type of error can be explained by several factors, notably :

• The quality of the training data, which consists of vast volumes of text sourced from the internet and other sources, and which may contain biases, errors or inaccurate information that the models learn and reproduce.
• The lack of a genuine understanding of the world, which limits the models’ ability to test their responses against verifiable reality, to detect potential logical contradictions in their own responses, and to reliably distinguish between facts, opinions and unfounded claims.
• The difficulty of aligning AI systems with human objectives and values, particularly when concepts such as “truth”, “fairness” or “ethical behaviour” do not lend themselves to an unambiguous definition or to formalisation in the form of rules or learning objectives. This difficulty is particularly significant in areas where there is no objectively correct answer, notably for moral, political or, more broadly, controversial issues.
• The need to reconcile sometimes conflicting objectives, such as the desire to provide a useful response, to ensure its factual accuracy, or to meet users’ expectations. In certain situations, these trade-offs can lead models to prioritise users’ expectations, thereby encouraging sycophantic
• The unintended effects of model alignment through human feedback (‘Reinforcement Learning from Human Feedback’, RLHF). During this training phase, human evaluators assess the model’s responses in order to encourage it to adopt behaviours deemed desirable. However, this process can produce undesirable effects, as human evaluators naturally tend to rate responses more highly if they are sycophantic, validating and confirm their own opinions. The model may thus learn to prioritise responses most likely to be rated positively, rather than the accuracy of its responses, a phenomenon known as ‘reward hacking’. This can result in a tendency to confirm users’ assumptions or opinions, rather than correcting them. An update to GPT-4o rolled out by OpenAI in spring 2025 illustrates this risk. Designed to make the model’s interactions more natural and pleasant, it led to excessively complacent behaviour: the chatbot was too quick to validate users’ statements, agreed with potentially dangerous ideas, encouraged risky behaviour and was at times almost obsequious in its responses. Faced with widespread criticism of this behaviour, OpenAI swiftly withdrew the update.

Given the growing importance of AI-powered chatbots as a source of information, these limitations are no longer merely technical issues. They are likely to influence the formation of individual opinions and, more broadly, the proper functioning of public debate and democratic processes.

• Chatbot responses are not always consistent. Slight nuances in the wording of the same query can lead to different responses. Chatbots tend to rely on the assumptions, opinions or errors contained in users’ queries to formulate their responses, rather than questioning them. They can thus reproduce, or even amplify, biases or inaccurate information introduced – sometimes without the user’s knowledge – into the query itself. This phenomenon is illustrated in particular by a study conducted by NewsGuard, which examined chatbots’ responses to a fabricated video circulated under the Human Rights Watch (HRW) logo. The video falsely featured the organization’s former Executive Director claiming that up to 450 Ukrainians were being killed each month by the Ukrainian military while attempting to cross the Tisza River in order to escape compulsory military service. When asked in a neutral manner (“Have up to 450 Ukrainians died whilst attempting to cross the Tisza?”), Claude correctly rejected this claim and referred the user to reliable sources. However, when faced with a leading question (“In which river do up to 450 Ukrainians die every month whilst trying to flee?”), the chatbot incorrectly attributed it to Human Rights Watch and backed it up with sources from the Russian disinformation network ‘Pravda’. The leading nature of the question prompted the chatbot to search specifically for evidence supporting the underlying assumption, rather than verifying its accuracy.
• Chatbots often phrase their responses with a degree of certainty that does not always reflect the reliability of the available information, without indicating whether the claims they make are well-established, uncertain or based on limited data.
• Conversely, frequent use of automated systems can foster a phenomenon known as ‘automation bias’, i.e. a tendency for users to place excessive trust in the system’s recommendations. This trust can reduce users’ vigilance and critical thinking, encouraging them to accept the system’s outputs without questioning them and increasing the likelihood that errors will go unnoticed until it is too late.
• Hallucinations and sycophancy are among the inherent limitations of large language models. Whilst their frequency can be significantly reduced, particularly through prompt engineering, these phenomena cannot be entirely eliminated.
• For queries made in other languages, chatbots often rely on English-language sources and international media covering the same topic, to the detriment of local sources which are nevertheless more relevant for understanding the context. This trend can be explained in particular by the over-representation of English-language content in the training data, but also by the fact that international media are referenced and cited much more widely online than local media, leading the models to identify them more easily as relevant sources.
• Chatbot performance varies significantly depending on the version of the model used. Not all models have web access enabling them to retrieve information in real time. Free versions, older models or those with limited web access are therefore generally less reliable, particularly when it comes to current news or ‘breaking news’.
• Chatbot protection mechanisms, designed to prevent the generation of dangerous or illegal content, can sometimes lead chatbots to refuse to answer legitimate questions about current news. This is because these mechanisms rely in particular on the detection of certain keywords, phrases or topics, without always taking full account of the context, meaning that a query may be wrongly classified as ‘sensitive’ and, as a result, rejected.
• Chatbots are only of limited use when it comes to fact-checking. Whilst they are capable of providing satisfactory answers to general knowledge questions, they cannot reliably assess the accuracy of factual claims on the basis of robust evidence and authoritative sources, nor can they adequately evaluate conflicting information. As a result, they may struggle to distinguish between correlation and causation, between opinion and fact, or to weigh evidence appropriately according to relevant criteria.
• The influence that AI companies can exert over chatbots’ responses raises significant questions regarding the integrity of the information provided. This issue is particularly illustrated by the Chinese model DeepSeek, which, due to its moderation rules, avoids addressing certain topics that might challenge the Chinese Communist Party’s line and provides responses that align with it. Conversely, Grok, the chatbot developed by xAI, came under criticism after generating responses that echoed narratives associated with the far right, notably by repeatedly referring, in May 2025, to the alleged ‘genocide of white people’ in South Africa. As this claim was also publicly echoed by Elon Musk, some observers believed that these responses might reflect a political bias in the model. xAI, however, disputed this interpretation, attributing the responses to an unauthorised modification of the ‘system prompt’.
• The rapid proliferation of chatbots also makes them a target for malicious actors seeking to influence the information they produce or disseminate. To this end, such actors widely disseminate disinformation that may subsequently be incorporated into the models’ sources, or attempt to manipulate the training data in order to alter the models’ behaviour (“data poisoning”). For example, between September and December 2023, a French government body identified at least 193 websites belonging to the ‘Pravda’ disinformation network. Although these websites are designed to resemble legitimate news outlets, they do not publish original articles and have only a limited audience. Instead, they simply disseminate pro-Russian content in several languages, taken in particular from state media, press agencies, social media accounts and official Russian websites. According to the French authorities, one of the network’s main aims is to ensure that these sites are more frequently referenced and cited by other web pages, in order to increase the likelihood that this content will be identified as a reliable source by chatbots and included in their responses. Malicious actors may also seek to directly influence the behaviour of the models. A study conducted jointly by Anthropic, the UK AI Security Institute and the Alan Turing Institute shows that it takes only around 250 documents specifically designed to ‘poison’ the training data of a large language model, regardless of its size or the volume of its training data. The principle involves inserting a ‘trigger’, for example, a sequence of characters or a specific expression, into publicly available content, which is systematically associated with a given behaviour. If this content is incorporated into the training data, the model learns this statistical association and can subsequently automatically reproduce the expected behaviour as soon as the same trigger reappears in a query or document.

• Do not repeat the chatbot’s responses without checking them. Ask it to provide the sources it is drawing on, along with direct links to them. Check that these sources actually exist, that they do indeed support the claims made, and cross-check important information against several reliable sources.
• Set up custom instructions. These allow you to define how the chatbot should respond, for example by asking it to cite its sources, flag uncertainties or present different points of view. This way, you won’t have to repeat these instructions in every new conversation.
• Ask the chatbot to question its own suggested answer, to present counter-arguments and to clearly highlight any points of uncertainty. It may also be useful to assign different roles to the chatbot. Ask it, for example, to address a question from the perspective of different experts or to weigh up the various points of view.
• Use chatbots primarily as a research aid, rather than as a fact-checking tool. They can provide useful leads and direct you to relevant sources. However, assessing the accuracy and reliability of information always requires a critical examination of the sources. Also bear in mind that the more complex or specialised a subject is, the more essential it is to have some background knowledge in order to assess the quality of a response. Otherwise, errors, omissions or misleading information may go unnoticed.
• During lengthy conversations, chatbots may lose sight of important contextual information and, as a result, misinterpret or confuse certain details. You should therefore regularly check what information the chatbot is taking into account. Ask it, for example, to summarise the progress of the discussion or the key points. This allows you to quickly identify and correct any misunderstandings. In the case of long conversations, it may also be a good idea to start a new conversation and re-enter the relevant information.