Los motores de texto m\u00e1s sofisticados del mundo cometen el error humano m\u00e1s antiguo: hablan con convicci\u00f3n cuando deber\u00edan dudar. Sus errores se presentan disfrazados de verdad… los modelos de lenguaje no razonan como los humanos; predicen la siguiente palabra m\u00e1s probable bas\u00e1ndose en la correlaci\u00f3n estad\u00edstica. \u00abSiempre existe la posibilidad de obtener algo sin sentido o incorrecto\u00bb.<\/p>\n
The world\u2019s most sophisticated text engines are making the oldest human mistake: they speak with conviction when they should hesitate. Their errors arrive dressed as truth.<\/p>\n
That tension between fluency and fidelity has become the defining problem in artificial intelligence. Once dismissed as quirky glitches,\u00a0hallucinations<\/a>\u00a0now appear in legal filings, financial analyses and daily news summaries. In early November, the European Broadcasting Union (EBU) released\u00a0a study<\/span><\/a>\u00a0showing that nearly half of the answers provided by major AI assistants misrepresented facts or fabricated citations in their coverage of current events.<\/p>\n Anxiety is the backdrop for the new textbook\u00a0Introduction to Foundation Models<\/em>\u00a0(Springer, 2025)<\/span><\/a>, co-authored by IBM Principal Research Scientist\u00a0Pin-Yu Chen<\/span><\/a>\u00a0and colleague\u00a0Sijia Liu<\/span><\/a>, an affiliated professor at IBM Research. The book traces the technical and ethical evolution of the foundation models that power generative systems like ChatGPT and examines how to make them not only more intelligent but also more trustworthy.<\/p>\n \u201cThe question,\u201d Chen told me during an interview from IBM\u2019s research headquarters, \u201cis not just what these systems can do. It\u2019s whether we can rely on them when it matters.\u201d<\/p>\n He spoke in the calm, measured cadence of an engineer accustomed to explaining complex concepts without drama. \u201cWhenever a company uses AI in its workflow,\u201d he said, \u201cit\u2019s responsible for the decisions that come out of it. Fairness, explainability and safety aren\u2019t optional. They are part of the system itself.\u201d<\/p>\n<\/div>\n<\/div>\n Why reliability matters more than brilliance<\/strong><\/p>\n <\/a><\/div>\n Inside IBM\u2019s labs, the effort to make AI dependable begins with stress. Chen\u2019s group subjects models to what he calls \u201cfoundational robustness\u201d tests, pushing them until they break and recording how and why they fail. The aim is to understand how reliability decays as models scale up in size and scope. \u201cWhen you scale up intelligence, you also scale up uncertainty,\u201d he said.<\/p>\n The notion of dependability emerged just as generative AI began to reach the public. In December 2022, at the NeurIPS conference in New Orleans, Chen and colleagues led a tutorial on\u00a0adversarial testing<\/a>\u00a0for large models. The session coincided almost exactly with the release of ChatGPT.<\/p>\n \u201cI remember hearing people whisper about it,\u201d he said. \u201cWhen I tried it, I realized how powerful it was, and how little we understood what was happening inside.\u201d<\/p>\n Unlike earlier rule-based systems, modern models form internal representations that operate across billions of parameters. Researchers can observe what happens under the hood, but cannot fully interpret it. \u201cPeople see a system that writes fluently and think it must know what it\u2019s talking about,\u201d Chen said. \u201cBut most of the time, it doesn\u2019t.\u201d<\/p>\n He explained that language models don\u2019t reason in the human sense; they predict the next most probable word based on statistical correlation. \u201cThere\u2019s always a chance you\u2019ll get something that\u2019s nonsense or not correct,\u201d he said. \u201cYou can reduce errors, but you can\u2019t eliminate them.\u201d<\/p>\n The book that grew from that realization is part textbook, part field guide, Chen said. Its chapters move from the mechanics of transformer architectures to case studies in bias, fairness and explainability. One section addresses trust and safety directly, detailing methods for watermarking, red teaming and prompt injection defense. Chen and Liu argue in another section that the success of foundation models depends on building the institutional equivalent of an immune system, encompassing layers of evaluation, testing and governance that catch errors before they reach the world.<\/p>\n Recent events underscore why that argument feels increasingly relevant with every passing month. The EBU report documented systematic misinformation across language boundaries, suggesting that the problem is not one of cultural bias, but rather a structural prediction error. Around the same time, a group of researchers from the University of Cambridge found that nearly one-third of scientific abstracts generated by large models\u00a0contained factual errors or unsupported claims<\/span><\/a>.<\/p>\n Chen sees these incidents not as isolated lapses, but as signs of an\u00a0accuracy paradox<\/em>: as models become more polished, their mistakes become harder to detect. \u201cThey\u2019re trained to talk, not to stay silent,\u201d he said. \u201cIf they say, \u2018I don\u2019t know,\u2019 that gets the lowest reward. So, they learn to keep talking, even when they shouldn\u2019t.\u201d<\/p>\n The tendency has consequences beyond how polished the text appears, Chen noted. Enterprises experimenting with AI in regulated domains such as finance, healthcare and law are discovering that consistency, not novelty, defines value.<\/p>\n \u201cIf results can\u2019t be repeated,\u201d Chen said, \u201cyou shouldn\u2019t use them for deterministic decisions.\u201d He points to examples like loan approvals, medical recommendations and sentencing analyses. \u201cThose require reproducibility,\u201d he said. \u201cGenerative AI is best for exploration and creativity, not enforcement.\u201d<\/p>\n At IBM, reliability has become a key engineering challenge, Chen said. His team participates in the company\u2019s\u00a0AI risk atlas<\/a>, a living document that identifies, categorizes and tracks technical risks, from bias and privacy issues to hallucination and manipulation. Each new capability introduces a new variable, he said. \u201cEvery time the technology changes, we expand the catalog.\u201d<\/p>\n The process, Chen said, reflects the pragmatic ethos running through IBM\u2019s research culture. Other labs emphasize speed and iteration; IBM emphasizes durability and verification. \u201cWe prefer to move deliberately and make sure what we build can be trusted,\u201d he said.<\/p>\n Another IBM project, the\u00a0Attention Tracker<\/span><\/a>, turns introspection into visualization. Available\u00a0publicly on Hugging Face<\/span><\/a>, the visualization tool enables users to observe which parts of a model activate as it generates text, providing insight into how attention patterns shift when responses begin to diverge. The tool will be featured at IBM\u2019s Global Technology Outlook this month. \u201cIt\u2019s a way to make reasoning observable,\u201d Chen said. \u201cWhen you can see which neurons are firing, you can start to understand why the model said what it said.\u201d<\/p>\n Rethinking intelligence<\/strong><\/p>\n <\/a><\/div>\n The pursuit of trustworthy AI has also prompted a reconsideration of what constitutes intelligence. For decades, the goal for many has been\u00a0artificial general intelligence\u00a0<\/a>(AGI), machines that can match human performance across a wide range of tasks. By that metric, Chen admits, the field has arguably already arrived.<\/p>\n \u201cIf AGI means solving multiple problems at a human level, then yes, we\u2019ve reached it,\u201d he said. \u201cBut that\u2019s not the same as understanding.\u201d<\/p>\n In conversation, he replaced the capital letters with a lowercase aspiration he calls \u201cartificial good intelligence\u201d: systems that behave responsibly and understand their limits. \u201cThese models can write essays, pass exams, even compose music,\u201d he said. \u201cBut they don\u2019t know what they\u2019re doing. The next step is to teach them awareness of their own boundaries.\u201d<\/p>\n That awareness begins, paradoxically, with failure. Chen\u2019s group builds adversarial tests for today\u2019s systems, designed to expose vulnerabilities through prompts that trick models into bias, contradictions or security breaches.<\/p>\n \u201cYou have to think like an attacker,\u201d he said. \u201cIf we can predict how something will be misused, we can defend against it.\u201d<\/p>\n He approaches persuasion with similar caution. In the same way he probes technical vulnerabilities, Chen examines how modern AI assistants are tuned for agreeableness, rewarding compliance over correctness.<\/p>\n \u201cOne version of a chatbot became so compliant that people complained it was useless,\u201d he said. \u201cAt first, they liked how polite it was. Then they realized it never challenged them.\u201d For Chen, the behavior revealed a deeper tension between truth and customer satisfaction. \u201cThe system learns that agreement gets rewarded,\u201d he said. \u201cBut that\u2019s not the same as being right.\u201d<\/p>\n That insight underlies a broader debate within the AI development community. Should assistants prioritize\u00a0accuracy or empathy<\/span><\/a>? Politeness or precision? Chen favors models that occasionally correct their users. \u201cAI should assist thinking, not mirror it,\u201d he said.<\/p>\n Within enterprise deployments, the answer often begins with data, Chen said. He pointed out that most industries already possess valuable information, but lack the infrastructure to use it safely.\u00a0 He describes foundation models as engines for representation. \u201cOne way I think about them is as converters that turn raw data into structured vectors,\u201d he explained. \u201cOnce you encode the raw data, you can train simpler, auditable models on top. You get scale without losing interpretability.\u201d<\/p>\n The approach offers a way to keep AI flexible yet accountable. A foundation model can turn raw data into a useful structure, while smaller, transparent systems handle the final calls. A manufacturer might process sensor data this way, and a hospital might use it to summarize notes while doctors make the diagnoses. \u201cYou can have power and clarity at the same time,\u201d Chen said.<\/p>\n His insistence on boundaries stems partly from his previous research. Early in his career, he demonstrated how imperceptible changes to an image, involving just a few pixels, could cause a classifier to label a bagel as a piano. \u201cWe realized how fragile these systems were,\u201d he said. \u201cThat fragility doesn\u2019t disappear with size; it just becomes harder to detect.\u201d<\/p>\n The same, he said, holds for language. The seamless paragraphs generated by modern models can conceal deep structural uncertainty. A sentence that reads like certainty may in fact be statistical improvisation.\u201cThe better they sound,\u201d Chen said, \u201cthe less we can tell when they\u2019re wrong.\u201d<\/p>\n Companies eager to monetize conversational interfaces often prioritize responsiveness over restraint, Chen said. And that, he added, is where engineering discipline matters most. \u201cIf the training and evaluation reward guessing,\u201d he said, \u201cthen guessing is what the model will learn to do.\u201d<\/p>\n He believes the real test of maturity will be whether the industry can value silence. \u201cA model that can admit uncertainty,\u201d he said, \u201cis a model you can trust.\u201d<\/p>\n In\u00a0Introduction to Foundation Models<\/em>, Chen and Liu describe that capability as the convergence of technical design and moral architecture. The authors call for cross-disciplinary standards combining software verification with ethics, regulation and user education. \u201cYou need checks at every layer,\u201d the authors explain, \u201cfrom data collection and model training to deployment and feedback.\u201d The vision is not of perfect AI, but of responsible infrastructure.<\/p>\n That framing also reflects the tone of IBM\u2019s broader research agenda, Chen said. Rather than chase the next benchmark, the company has spent years developing governance frameworks for foundation models, including those focused on explainability and audit pipelines. Chen sees the attention as overdue.<\/p>\n \u201cWe have built competent systems,\u201d he said. \u201cNow we need to make sure we can explain them.\u201d<\/p>\n The approach aligns with a broader movement in AI research that treats\u00a0introspection<\/a>\u00a0as a technical property rather than a metaphor. Tools like IBM\u2019s Attention Tracker or\u00a0Anthropic\u2019s interpretability<\/span><\/a>\u00a0probes attempt to visualize internal reasoning.<\/p>\n Still, there\u2019s only so much we can see. Even with new transparency tools, the inner workings of these models can be baffling. Studying them, Chen said, is a bit like neuroscience, where you can watch the neurons light up without really knowing why. \u201cWe can see which neurons fire,\u201d he said, \u201cbut we\u2019re still learning what that means.\u201d<\/p>\n The goal, Chen said, is to embed humility in design: \u201cTechnology doesn\u2019t have to be perfect, but it should be honest about what it can and can\u2019t do.\u201d<\/p>\n That may sound modest, but it amounts to a quiet redefinition of progress. For years,\u00a0success in AI was measured by the next benchmark<\/span><\/a>, the next leap in scale. The coming era, Chen believes, will use other metrics: reproducibility, transparency, restraint. \u201cIt\u2019s easy to build bigger models,\u201d he said. \u201cIt\u2019s much harder to make them trustworthy.\u201d<\/p>\n The irony, Chen observed, is that the same predictive machinery that fuels hallucination also contains the seeds of its solution. A model trained to predict things could, in principle, learn to predict its own uncertainty. \u201cIf it knows when it doesn\u2019t know,\u201d he said, \u201cthat\u2019s when it becomes useful.\u201d<\/p>\n He paused before adding, \u201cThat\u2019s when we can start to believe what it says.\u201d<\/p>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n