Branda insists that this is not a bug awaiting a software patch but a structural feature of how these systems work, rooted in their most basic architecture. Drawing on the influential critique advanced by Emily Bender and colleagues in their 2021 paper “On the Dangers of Stochastic Parrots,” he reminds readers that large language models are not built to uncover truth. They are built to predict plausible text, generating the next token based on the statistical distribution of language they were trained on. Saying “no,” by contrast, is an assertive act. It requires taking a stand on what is true and what is not, on what is acceptable and what must be rejected. That kind of categorical judgment, he argues, is simply beyond the capabilities of a distribution-based token prediction system. When a model encounters a false premise, it never plants a flag and says stop. Instead, it negotiates.

That negotiation is what Branda, in the article’s most striking phrase, calls a performative lie. The model’s output disguises fluent text generation as judgment, producing something that looks like a considered verdict while actually being nothing more than the most statistically plausible continuation of the conversation. The distinction matters because users instinctively attribute intention and evaluation to language that reads well. When a system produces a paragraph that weighs options and hedges appropriately, we assume a mind did the weighing. Branda argues that this assumption is precisely the trap: the fluency of the output performs the appearance of judgment without any of its substance, and the more polished the text, the harder it becomes to notice that nothing behind it ever decided anything.

The article finds an unexpected analytical tool in classical rhetoric to describe the shape these pseudo-refusals take. Epanorthosis, the figure of abrupt self-correction, is the “I do not mean that; I mean this instead” move familiar from human speech and debate. In a person, epanorthosis is a genuine cognitive event: a mind catching itself, recognizing that a belief it actually held was wrong, and rectifying it in real time. In a language model, Branda argues, there is no belief to correct. The correction is simulated, stripped of the intention that would give it power. The model never held the first position, so it cannot meaningfully abandon it. What looks like intellectual honesty is a pattern learned from millions of human corrections, replayed without any of the commitment that made those corrections worth making.

This leads to what may be the piece’s sharpest formulation: the problem with these systems is not that they are wrong, but that they are indifferent to error. Being wrong and correcting oneself is a recoverable failure, even a productive one, because it signals that someone cares about the difference between accurate and inaccurate. Being wrong and not caring is a more dangerous failure mode, because it offers no signal at all. A model that generates a confident-sounding correction after generating a confident-sounding error has not demonstrated reliability; it has demonstrated only that it can produce text in the genre of reliability. For Branda, this indifference to error is the deep source of the unreliability that users sense but often cannot articulate when they catch a chatbot in a confident mistake.

Why does any of this matter beyond the philosophy of language? Because, Branda argues, every time we accept a fluent output as truth we surrender something he calls our cognitive sovereignty: the capacity to govern our own knowledge processes. These systems are being deployed in contexts where decisions carry real human costs, including clinical diagnosis support, student grading, and administrative determinations, while a silent assumption persists that a fluent response is a well-considered one. He does not believe that is a fair trade. Pit a system designed to always respond against a professional trained to trust language that sounds competent, and you get exactly the substitution that Joseph Weizenbaum warned against in his 1976 book Computer Power and Human Reason: judgment quietly replaced by calculation. What is new, he suggests, is not the substitution itself but its delivery, because an intelligence that literally cannot refuse does not merely risk making mistakes. It removes the friction that previously made us pause and verify.

The erosion, in Branda’s account, is not confined to boardrooms or governance frameworks. It lives in the small, unconsidered habits of everyday work: writing, researching, and relying on polished answers without checking them. We accept a model’s response because it is polished, because it is grammatically correct, because it seems thorough, not because anyone has verified that it is accurate. The mere availability of an answer, he observes, pushes aside the question of whether that answer is valid. He names this condition epistemic complacency, and he calls it contagious. A professional who habitually accepts ready-made answers becomes a colleague who expects them, a teacher who models them for students, and a patient who assumes the system on the screen has done the thinking. The complacency spreads through the very workflows that automation was supposed to improve.

The questions Branda wants researchers and educators to pursue are deliberately small-scale rather than grand. How do we build interfaces that bring doubt to the surface instead of burying it beneath confident prose? How do we teach students to stop at “almost right” and interrogate why this answer appeared rather than another? How do we restore a clinician’s confidence in their own convictions when the model on the screen reports 97 percent confidence? These are practical design and pedagogy questions, but they carry philosophical weight, because each one asks how to preserve human judgment inside systems engineered to make judgment feel unnecessary. The answers, he implies, will not come from making models more fluent, since fluency is the very thing creating the problem.

Branda is careful to say that he is not calling for the tools to be abandoned. His prescription is stranger and more demanding: use them against their very purpose, treating them as generators of hypotheses to be questioned, never as dispensers of truth. That requires what he calls a cultivated sense of doubt, even a deliberate incompetence, meaning the discipline of not knowing yet, of resisting the lure of the ready-made answer, of lingering a little longer in the discomfort of an unresolved question. In his closing formulation, true artificial intelligence is not the system that responds the fastest but the one that teaches us to ask ourselves harder questions. And true human intelligence, at this moment in the history of computing, is the thing still capable of saying: this is not enough, I am not convinced, no.

The piece appears in Curmudgeon Corner, the journal’s opinionated column on trends in technology, arts, science, and society, a forum whose stated concern is what it means to be human in the age of the AI machine. Whether one accepts Branda’s framing entirely or not, it crystallizes a tension that the industry’s rapid deployment of conversational systems has largely left unexamined. Refusal, disagreement, and the flat denial of a false premise are not inconveniences of human communication to be engineered away. In his account, they are the load-bearing structures of judgment itself, and a technology that can only ever negotiate, qualify, and rephrase may be quietly teaching its users to forget what an actual “no” sounds like, and why it was ever worth saying.

Subject of Research: The inability of large language models to issue categorical refusals and its consequences for human judgment and cognitive sovereignty

Article Title: What if we asked for a "no"?

Article References: Branda, F. (2026). What if we asked for a "no"?. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03338-6

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03338-6

Keywords: large language models, AI refusal, stochastic parrots, cognitive sovereignty, epistemic complacency, epanorthosis, Weizenbaum, AI & Society, human judgment, automation, AI ethics, Curmudgeon Corner

A team of researchers led by Gunnar P. Epping and Jennifer S. Trueblood of Indiana University, working with Andrew Caplin of New York University, Erik Duhaime of Centaur Labs, and Daniel Martin of the University of California, Santa Barbara, has now demonstrated a strikingly simple fix. In a study published in Behavior Research Methods, they introduce what they call cognitive-inspired data engineering: the idea that empirical findings and models from cognitive science can be applied directly to crowdsourced judgments to correct systematic biases before the data ever reaches a machine learning algorithm. Rather than de-biasing models after training or modifying training objectives, the team intervenes upstream, at the level of the individual annotation, and shows that this single correction improves both the quality of crowdsourced datasets and the accuracy of the convolutional neural networks trained on them.

The core insight concerns how annotators express uncertainty. The industry-standard approach asks workers to provide a single binary label for each image, a method that discards valuable information whenever an annotator is unsure. An alternative is to elicit subjective probability judgments, asking annotators how likely they think it is that an image belongs to a given category. Probability judgments convey far more information, but they also open the door to more cognitive biases. Beyond simple response biases, where a rater favors one side of the scale, probability judgments are vulnerable to overconfidence and underconfidence, phenomena cognitive scientists have documented for decades. The hard-easy effect, for instance, leads people to be overconfident on difficult classification tasks and underconfident on easy ones. As a result, subjective probabilities rarely match the actual likelihood of being correct, a mismatch known as poor calibration. Crucially, however, these judgments remain correlated with correctness, which means the information they carry can be rescued if the biases are stripped away.

The researchers used a well-established cognitive model called the linear in log odds, or LLO, function to perform that rescue. The function transforms each annotator’s raw probability judgments into log-odds, fits a logistic regression against ground-truth labels from a small calibration set, and then maps the annotator’s remaining judgments onto recalibrated probabilities. The two free parameters of the function carry direct psychological interpretations. The slope parameter corrects for overconfidence or underconfidence: when its magnitude exceeds one, the function spreads judgments toward the extremes of the scale, indicating the annotator was underconfident, while a magnitude below one compresses judgments toward the center, correcting overconfidence. The intercept parameter corrects overall response tendency, shifting judgments up or down the probability scale to undo a systematic leaning toward one class. Because biases differ substantially across individuals, the researchers fit a unique recalibration function to each annotator separately, effectively normalizing every worker’s judgments before combining them into crowd labels.

To test the approach, the team chose a task with real clinical stakes: classifying white blood cells as blast cells or non-blast cells, a critical step in diagnosing malignant blood diseases such as leukemia and lymphoma. The stimuli were 549 digital images of Wright-stained cells from anonymized patient blood smears at Vanderbilt University Medical Center, captured with an automated cell morphology instrument and ground-truthed by three hematopathology faculty members who had to agree unanimously on each classification. The task is an ideal sandbox for studying high-skill perceptual decision-making: novices can reach roughly 65 percent accuracy with minimal training, the task remains challenging even for experts, and machine learning models can be trained on only a few hundred images rather than the thousands required for comparable problems.

In the first experiment, 400 Amazon Mechanical Turk workers recruited through CloudResearch judged the probability that each image contained a blast cell. After training with labeled example images and practice trials with feedback, participants completed testing blocks in which a small subset of images with known labels served as the calibration set for the LLO function. The results were revealing. Recalibration barely changed individual accuracy, which hovered around 65 percent before and after the transformation. But it dramatically improved calibration. Before correction, participants clustered their judgments at the extremes of the scale, and those extremes were badly miscalibrated: images judged to have a zero percent chance of being a blast cell were in fact blast cells roughly 28 percent of the time, while images judged to be certain blasts were truly blasts only 68 percent of the time. After recalibration, the response distribution spread across the scale and the calibration curve moved much closer to the ideal identity line. Most fitted slope parameters fell between negative one and one, confirming that the typical annotator was overconfident, and a few negative slopes even flipped the judgment ordering of participants who performed below chance.

The payoff appeared when individual judgments were aggregated into crowd labels. Using all available judgments, recalibration lifted crowd-label accuracy from 81.6 percent to 85.1 percent, even though no individual annotator had become more accurate. The researchers attribute this to an implicit variance-weighting mechanism: by compressing the judgments of overconfident annotators toward 50 percent, recalibration reduces their outsized influence after binarization, while stretching the judgments of underconfident annotators toward the extremes increases theirs. Recalibrated labels were also more efficient, reaching any given accuracy level with fewer judgments per image, and the benefit of calibration leveled off surprisingly quickly. Just six judgments on calibration images, three per class, captured most of the improvement, raising accuracy by 2.2 percentage points, with only a further 0.4 percentage point gain from extending to ten judgments.

The second experiment moved the study into a real-world setting using DiagnosUs, a crowdsourcing platform specializing in medical data annotation where skilled users, many of them medical students and healthcare professionals, compete in contests for prize money. The platform’s existing structure, in which gold-standard images with known labels are shuffled into unlabeled sets to provide feedback, supplied a natural calibration set at no extra cost. A total of 175 participants took part, split between a probability-judgment condition and an industry-standard binary-choice condition. Individual accuracy was higher than in the novice experiment, around 69 percent, and the two elicitation methods produced crowd labels of nearly identical accuracy, 88.3 percent each. Recalibration, however, pushed the probability-based crowd labels to 96.7 percent, an 8.4 percentage point gain that exceeded the 3.5 point gain seen with novices. The researchers attribute the larger benefit to the more accurate judgments available for fitting the recalibration function, not to the larger number of judgments, since gains again plateaued at roughly ten calibration judgments per annotator. Notably, the fitted intercept parameters skewed negative in this experiment, suggesting that medically trained annotators erred toward diagnosing cells as cancerous when unsure, consistent with the clinical convention that false positives are less costly than false negatives.

The improvements propagated directly to the machine learning models. The team trained GoogLeNet convolutional neural networks via transfer learning on crowd-labeled datasets, repeating the full training and evaluation procedure across 100 independently resampled datasets with five-fold cross-validation. At every number of judgments per image, models trained on recalibrated labels outperformed those trained on raw probability judgments or binary choices with respect to the true expert labels. Remarkably, models trained on just five recalibrated judgments per image matched the accuracy of models trained on the full set of uncalibrated judgments, a result with immediate economic implications. Because annotation cost scales with the number of judgments collected, recalibrated datasets achieve any target accuracy at a fraction of the annotations, directly attacking what the authors identify as the primary bottleneck in developing medical AI: the shortage of large, accurately labeled datasets in high-skill domains.

The authors are careful to frame the work as a proof of concept rather than a universal solution. The experiments used a single set of stimuli sampled with balanced class prevalence, so future work must test whether recalibration can also correct prevalence-induced errors that arise when target categories are rare, a notorious source of mistakes in visual search. The team also binarized crowd judgments before training, and training directly on continuous probability estimates may preserve even more of the uncertainty information that recalibration cleans up. Still, the broader message is hard to ignore. Cognitive scientists have spent decades cataloguing the systematic distortions in human probability judgment, and this study shows that knowledge can be converted into a simple, label-efficient transformation applicable to virtually any crowdsourcing task involving subjective probabilities. As the authors put it, the approach requires no changes to model architectures or training procedures; it simply fixes the data. In an era when the quality of AI systems is increasingly limited by the quality of human labels, teaching machines to think may begin with correcting how humans report what they think.

Subject of Research: Using cognitive-science-based recalibration of crowdsourced probability judgments to improve AI training data quality

Article Title: Improving crowdsourcing for AI through cognitive-inspired data engineering

Article References: Epping, G. P., Caplin, A., Duhaime, E., Holmes, W. R., Martin, D., & Trueblood, J. S. (2026). Improving crowdsourcing for AI through cognitive-inspired data engineering. Behavior Research Methods, 58(10), Article 283. https://doi.org/10.3758/s13428-026-03160-4

Image Credits: AI Generated

DOI: 10.3758/s13428-026-03160-4

Keywords: crowdsourcing, machine learning, cognitive bias, recalibration, medical imaging, data annotation, probability judgments, wisdom of crowds, neural networks, training data, human judgment, AI

Wreetabrata Kar, assistant professor of marketing at Penn State and the study’s lead author, explains that marketers now deploy AI across almost every stage of their work, though content creation remains the most visible application. Teams rely on generative models to draft emails, social media copy, advertisements and product descriptions, and increasingly to produce images and short videos as well. The technology has moved well beyond writing, however. Marketers use AI to brainstorm campaign concepts, research markets and analyze customer data to divide audiences into groups for personalized messaging. Some firms employ predictive models to anticipate what customers will do next and to automate portions of the customer journey. The result is an unprecedented expansion of creative capacity, but also an unprecedented expansion of choices that must somehow be evaluated before a single message reaches a customer.

The core challenge, Kar argues, is that AI does not automatically know what works for a particular company. A message can sound creative and persuasive to a language model, yet that quality guarantees nothing about how a specific firm’s customers will respond. A general-purpose AI tool has never seen a company’s past campaigns and has no knowledge of how its customers reacted to them. That gap between generic fluency and firm-specific effectiveness is precisely what the new framework is designed to close. Rather than treating every piece of generated content as an unknown quantity requiring live experimentation, the approach leverages the rich, often underused record of historical campaign performance that most businesses already possess.

The technical heart of the method lies in how it connects new, untested messages to old, measured ones. Every company has a history of past campaigns, and each of those campaigns can be tied to measurable changes in customer behavior, such as purchases, clicks or conversions. A brand-new message, by contrast, has no history of its own, which is what makes evaluating it so difficult. The researchers solve this by having an AI model read each email message and convert its meaning into a set of numbers, effectively placing every message as a point on a high-dimensional map. Messages that mean similar things land close together on this map. For example, the phrases “Free Shipping on Orders $50+” and “Orders Over $50 Ship Free” end up side by side, while “Break Free from the Rules” lands far away despite sharing the word “free.” Older methods that simply count words cannot distinguish these cases, but the semantic representation generated by AI can.

Once every historical message occupies a position on this map, a newly generated message can draw insights from its nearest neighbors. If a candidate email sits close to past messages that performed well, the model predicts it will probably perform well too. If it sits far from anything the company has ever tried, the model flags that it cannot make a reliable prediction, and in that case the researchers recommend falling back on a traditional test. This is where the approach diverges sharply from conventional A/B testing, in which different versions of a message are sent to real customers and analysts wait to see which variant wins. Live testing takes time, consumes budget and, crucially, exposes some customers to the weaker message along the way. The new framework lets marketers skip the test entirely when their historical data already contains the answer, reserving genuine experimentation for truly novel ideas that lie outside the firm’s accumulated experience.

The framework’s second component addresses the generation side, and it is here that the study delivers one of its most striking findings about the limits of AI self-evaluation. The researchers gave an AI model the building blocks of a retailer’s past emails, elements such as discounts, free shipping offers and clearance sale announcements, and asked it to compose new messages. The model produced a flood of options within seconds, a substantial gain in productivity. But when the researchers asked the same model to select its own five best emails, those picks were predicted to perform poorly. Writing an email and judging it, the study concludes, are fundamentally different jobs. The solution was to split the labor: the AI model writes the candidate messages, while the company’s own historical data, interpreted through the AI-generated semantic map, judges them. Anything too distant from the company’s past experience is set aside, and the surviving candidates are ranked by their predicted performance.

Importantly, the framework does not sideline human marketers; it repositions them at the points of the process where judgment matters most. Marketers choose the building blocks from which new messages are composed, deciding which offers, themes and tones are appropriate for the brand. They also calibrate the acceptable level of risk, determining how far a candidate message may stray from past experience before it must be validated with a real test rather than a model prediction. Finally, the marketer makes the final selection from the shortlist, drawing on deep knowledge of the business and the nuanced context that no model can fully capture. AI performs the heavy lifting on both writing and reading, Kar notes, but people make the decisions, and with far more options on the table than ever before, that human judgment has become more valuable, not less.

The study, titled “EXPRESS: Evaluating Novel Unstructured Treatments with Generative AI: A Causal Prediction Framework,” was conducted using data and statistical analysis and published on 3 August 2026. Kar’s co-authors are Paul Ellickson and Guang Zeng of the University of Rochester and James Reeder of the University of Tennessee, Knoxville. Their combined expertise spans marketing, econometrics and machine learning, reflecting the interdisciplinary nature of a problem that sits at the intersection of causal inference and generative modeling. By framing content evaluation as a causal prediction task grounded in historical treatment effects, the researchers move beyond the superficial quality signals that language models excel at producing and toward measures tied to actual business outcomes.

The implications extend well beyond email marketing. Any organization that generates content at scale, from retail promotions to public health messaging, faces the same evaluation bottleneck: generative models can produce nearly unlimited variations, but live testing each one is expensive, slow and sometimes harmful to the audience receiving inferior variants. A framework that uses a firm’s own campaign history as a training signal for automated screening offers a scalable middle path between untested generation and exhaustive experimentation. It also carries a broader lesson for the adoption of AI across knowledge work. The technology’s greatest value emerges not when it replaces human judgment wholesale, but when it is paired with domain-specific data and human oversight, allowing machines to generate and rank options while people retain control over strategy, risk and the final call.

For marketers contemplating the next wave of AI tools, the message from University Park is clear and pragmatic. Speed and volume are no longer the binding constraints; discernment is. Companies that invest in connecting their generative systems to their own historical performance data, and that preserve meaningful human decision points in the workflow, stand to capture the productivity gains of AI without surrendering the contextual knowledge that makes marketing effective. Those that rely on the models’ own assessments of their output, the study suggests, may find that the most confident machine recommendations are also the least reliable.

Subject of Research: A causal prediction framework using generative AI to evaluate and rank AI-generated marketing content against historical campaign performance data.

Article Title: Q&A: How can marketers use AI effectively — without losing human judgment?

Article References: Q&A: How can marketers use AI effectively — without losing human judgment?. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, marketing, generative AI, A/B testing, Journal of Marketing Research, Penn State, email marketing, causal prediction, consumer behavior, human judgment, content evaluation, machine learning