The Seven Sins of AI in Education and Research

PERSPECTIVES · AI, EDUCATION & RESEARCH

Dandan Zhang · 5 November 2026

Figure 1. Seven habits that can weaken independent judgement in education and research. Generated by AI; conceptual illustration.

A student reports 96% accuracy for a model trained to distinguish soft from hard objects using tactile images. The number is correct for the evaluation they performed. In the AI-edited report, however, it becomes a claim that the model can distinguish unfamiliar objects reliably.

Imagine the supervisor asking how the test set was constructed. The student explains that the images were randomly divided between training and testing. The difficulty is that images of the same physical objects, sometimes acquired during the same contact sequence, appear in both sets. The model has not been tested on an object it has never encountered.

That is a hypothetical example, but the distinction is a real one in experimental design. An image-level split can evaluate performance on held-out images without establishing object-level generalisation. The error enters the scientific account when an AI-assisted edit silently turns a limited result into a stronger claim. The student's earlier note - "96% accuracy; not sure about new objects" - would have been more accurate.

I call the following problems the seven sins of AI in education and research. The names are deliberately pointed, but the habits themselves predate generative AI. A tool can make them easier to overlook by producing an apparently finished answer before anyone has examined the decisions behind it.

Figure 2. Testing repeated images of familiar objects does not establish performance on unfamiliar objects. Generated by AI; conceptual illustration.

1. Vanity: mistaking fluency for understanding

AI can produce clear explanations, well-structured arguments and technically convincing solutions. These capabilities are valuable in education, but they also make it increasingly difficult to distinguish between a student's ability to present an answer and their understanding of the reasoning behind it. A polished piece of work may conceal important gaps in knowledge that would otherwise become apparent during the learning process.

Consider a student learning backpropagation in a deep learning course. An AI assistant can explain the chain rule, derive the gradients and illustrate how information flows through a neural network. With this assistance, the student may prepare an excellent explanation of the algorithm. Yet when asked to consider a small network and predict how changing a particular weight would affect its output and loss, they may struggle to work through the problem independently.

The difficulty is not necessarily a lack of familiarity with the equations. Understanding backpropagation requires knowing how the quantities depend on one another, why the derivatives take their particular forms and how to recognise when a calculated result is implausible. These aspects of reasoning can be hidden when students encounter only a complete, polished derivation.

In teaching, I would encourage students to work through a simple example before consulting an AI-generated solution. They could predict the effect of changing a weight, calculate the relevant derivative and compare it with a numerical approximation. If the results disagree, the next task is to understand why. Was a dependency overlooked? Was the chain rule applied incorrectly? Does the prediction make sense given the structure of the network? Working through such questions can reveal more about a student's understanding than reproducing a correct derivation.

This does not mean that students should avoid AI-generated explanations. A well-designed AI tutor can make difficult concepts more accessible, explain the same principle in different ways and help students identify misconceptions. The concern arises when students accept a clear explanation as evidence that they have mastered the underlying concept, without attempting to apply it themselves.

The same distinction matters in research supervision. A sophisticated literature review or a convincing presentation may communicate an argument effectively, but students should also be able to explain the assumptions, evidence and decisions behind it. Recognising a convincing argument is not the same as being able to construct and defend one independently.

AI can improve how students explain what they know, but genuine understanding becomes visible when they can reason through unfamiliar problems themselves.

2. Dependence: fixing the script without diagnosing the measurement

One of the advantages of AI is that students can obtain immediate help with programming, mathematics and unfamiliar technical methods. It can remove frustrating obstacles and allow them to focus on more meaningful aspects of their work. However, if students routinely turn to AI before attempting to understand a difficulty, they may gradually lose opportunities to develop the diagnostic skills that independent research requires.

Consider a student working with tactile sensing data. They need to align video frames with force-sensor measurements, but the two arrays have different lengths. An AI assistant suggests truncating the longer array so that the data can be processed together. The code now runs, and the resulting plot appears reasonable. Yet the underlying problem may remain unresolved: the camera and force sensor could have started recording at different times or operated at different sampling rates. Matching their first recorded values does not necessarily mean that the measurements correspond to the same physical events.

A researcher needs to understand the acquisition process before deciding how the signals should be aligned. This might involve checking timestamps, identifying synchronisation events or testing the procedure using signals with a known time offset. AI can help implement the resulting method, but it cannot establish that the scientific assumptions are valid merely by producing executable code.

The educational concern is broader than programming. When a student receives an immediate solution to every difficulty, the final result may improve while the process of identifying, diagnosing and correcting mistakes receives less attention. Over time, this could make students more dependent on external assistance when encountering problems that do not have an obvious or established solution.

There is some relevant evidence from educational research. In a field experiment involving secondary-school mathematics, Bastani and colleagues [1] found that relatively unrestricted AI assistance improved students' performance during practice but was associated with poorer performance on a subsequent assessment without assistance. A more carefully structured AI tutor substantially reduced this problem. These findings should not be directly generalised to doctoral research, but they illustrate why completing a task with assistance and developing the ability to perform it independently need to be assessed separately.

I would not ask students to struggle unnecessarily with every technical obstacle. Learning to use appropriate tools is itself part of becoming an effective researcher. The important distinction is whether AI removes an incidental barrier or takes over the reasoning that the student needs to practise.

In supervision, I would encourage students to explain what they initially thought was wrong, what alternatives they considered and how they verified the solution. AI can then be used to explore further possibilities, generate tests or clarify unfamiliar concepts. The student remains responsible for understanding why the chosen approach works.

The purpose of AI assistance should be to strengthen students' problem-solving abilities, not to make independent problem-solving an activity they rarely practise.

3. Flattery: asking a tool to endorse a preferred conclusion

AI can generate persuasive explanations for ideas that have not yet been adequately tested. The difficulty is often not that the explanation is obviously incorrect, but that it accepts an assumption already embedded in the user's question. When researchers ask AI to justify a method they favour, they may receive a convincing argument without having critically examined whether the method actually delivers the claimed benefits.

Consider a student developing a robotic perception system that combines vision and touch. They ask an AI assistant, "Why is multimodal perception better than vision alone?" The question already presupposes that combining the two modalities produces an advantage. The assistant may explain that vision provides information about object geometry and position, while touch contributes information about contact and deformation. It may then suggest that combining these sources improves the robustness of robotic manipulation.

This is a reasonable account of why multimodal perception could be beneficial. It does not demonstrate that the student's particular system performs better, or establish why any observed improvement occurs.

Even if the combined model outperforms a vision-only baseline, further investigation is needed. Does tactile information make a genuine contribution? Is the improvement maintained under partial visual occlusion or changing contact conditions? Could the difference be explained by additional training data, increased model capacity or an unfair experimental comparison? Appropriate baselines and ablation experiments are necessary to distinguish these possibilities.

The same tendency can influence literature reviews. If a student asks AI to find evidence supporting the superiority of multimodal perception, it may assemble a persuasive account from studies reporting favourable results. Yet those studies may involve different datasets, tasks, model architectures or evaluation protocols. Findings that support a method under one set of conditions cannot automatically establish its advantages in another. Contradictory evidence and methodological limitations deserve attention alongside positive results.

This becomes particularly problematic during peer review. When a reviewer questions the contribution of a proposed method, an AI assistant may help construct an elaborate defence of the original claims. However, a more confident rebuttal cannot compensate for a missing baseline or an unsupported interpretation. Sometimes the scientifically appropriate response is to perform an additional experiment, reconsider the explanation or acknowledge a limitation.

I would encourage students to examine the assumptions behind the questions they ask AI. Instead of asking why their preferred approach is better, they could ask under which conditions it might offer an advantage, when it might fail and what evidence would distinguish its genuine contribution from alternative explanations. AI can be useful in generating these competing possibilities, but the researcher must evaluate them against the literature and experimental evidence.

Students should also feel able to challenge suggestions from their supervisors. An established research direction or a supervisor's preferred hypothesis should not be treated as correct simply because it comes from someone more experienced. The quality of a scientific argument depends on its reasoning and evidence, regardless of who proposes it.

AI should help researchers examine their assumptions, rather than simply provide more convincing reasons to believe what they already think.

Figure 3. A persuasive explanation still needs to be tested against observations, alternatives and experimental evidence. Generated by AI; conceptual illustration.

4. Conformity: choosing the fashionable method before identifying the problem

AI can help researchers explore unfamiliar methods and generate possible solutions. However, its suggestions may also encourage students to approach new problems through established techniques before they have properly examined the underlying scientific question. This is particularly concerning when an unexpected observation could lead to a genuinely new direction of research.

Consider a hypothetical microrobotics experiment in which a robot repeatedly deviates from its intended trajectory near a surface. A student asks an AI assistant how to improve its control, and the assistant recommends reinforcement learning. The suggestion sounds reasonable, especially given the success of learning-based control in many robotics applications. Yet it assumes that developing a more sophisticated controller is the appropriate response. The observed deviation might instead arise from near-surface physical interactions, inaccurate position estimation or limitations in actuation.

Before choosing a control method, I would encourage the student to investigate what is actually happening. Does the deviation change with distance from the surface? Can the robot's position be measured independently? How does its motion respond to controlled changes in actuation? These experiments may be less technically impressive than implementing a new learning algorithm, but they could reveal a physical mechanism or measurement limitation that deserves further investigation.

The concern extends beyond the selection of a particular algorithm. AI can readily suggest research directions based on familiar combinations of methods, applications and performance objectives. Such suggestions can be useful starting points, but students may be tempted to adopt them without considering whether they address the most interesting questions raised by their own observations. A project can acquire a convincing title and methodology before its scientific motivation has been properly established.

There is some indirect evidence worth considering. Doshi and Hauser [2] found that AI-generated suggestions improved evaluations of individual short stories while making the resulting stories more similar to one another. Their study concerns creative writing, and its findings cannot establish that AI is making scientific research more homogeneous. Nevertheless, it raises a relevant concern: improvements in the apparent quality of individual outputs do not necessarily preserve the diversity or originality of the ideas behind them.

I do not regard established methods as inherently less valuable than novel ones. A familiar technique may be exactly what a problem requires. The important question is whether it has been selected through scientific reasoning or simply because it was readily suggested. In supervision, I would encourage students to explain why a particular method is appropriate, which alternatives they considered and what observations might lead them to reconsider their choice.

Students should also have opportunities to pursue questions that emerge from their own experiments, even when these differ from the original research plan. AI can help them examine possible explanations and approaches, but it should not determine which questions are worth asking.

Scientific originality depends on our ability to recognise problems worth investigating, rather than simply selecting from solutions that are already familiar.

5. Excess: producing work packages faster than evidence

AI makes it remarkably easy to expand a research idea into an ambitious programme. A preliminary question can quickly become a proposal involving new algorithms, extensive datasets, multiple experiments and ambitious applications. The resulting plan may look comprehensive, but the ease of generating additional tasks can conceal how much work is actually necessary to establish a meaningful scientific contribution.

Consider a student proposing a twelve-week project to fabricate a tactile sensor, collect a large dataset, train several learning models and integrate the system into a robotic hand. An AI assistant could readily develop this into a detailed programme with objectives, milestones and expected contributions. Yet the entire plan may depend on a basic question that remains unanswered: can the proposed sensor produce measurements that are sufficiently reliable and repeatable for the intended task?

A preliminary calibration experiment using available equipment might therefore be more valuable than immediately pursuing the full programme. If the measurements vary unpredictably under controlled conditions, collecting additional data or developing a more sophisticated learning algorithm will not necessarily address the problem. The student may first need to improve the sensor design, reconsider the measurement procedure or redefine the research objective.

The difficulty is that an elaborate plan can create an impression of progress before the critical assumptions have been tested. Each additional experiment may appear reasonable when considered separately, while the overall project becomes increasingly difficult to justify or complete. AI can make this expansion particularly tempting because generating a new work package requires little effort compared with carrying it out.

This does not mean that students should avoid ambitious research. Ambition is valuable when accompanied by a clear understanding of priorities, dependencies and feasibility. I would encourage students to identify which uncertainties must be resolved first, what evidence would justify proceeding and which outcomes would require a change of direction. A well-designed project should help researchers make these decisions, rather than simply maximise the number of activities undertaken.

The same problem appears in academic writing. Asked to improve a manuscript, an AI assistant may recommend additional experiments, a broader literature review, more figures or substantial restructuring. Some suggestions may address genuine weaknesses. Others increase the length and complexity of the paper without strengthening its central argument. The author needs to distinguish between work that provides necessary evidence and work that merely makes the manuscript appear more comprehensive.

Research also requires the judgement to recognise when enough evidence has been gathered to support a carefully bounded conclusion, and when further investigation is genuinely necessary. An experiment that resolves a central uncertainty may contribute more than several additional comparisons with little bearing on the scientific question.

Good research is not defined by the volume of work proposed or completed, but by how effectively that work advances our understanding of the problem.

6. Inequity: equal access to a tool, unequal ability to challenge it

AI can make sophisticated analytical tools accessible to students with very different academic backgrounds. This is one of its advantages, particularly in interdisciplinary research. However, being able to generate an analysis is not the same as being equipped to evaluate it. Students with stronger foundations in statistics or experimental design may recognise questionable assumptions that others would not even know to look for.

Consider a hypothetical experiment involving ten tactile sensors, each measured a hundred times. An AI assistant processes the thousand measurements and produces a polished figure with error bars and statistical comparisons. The results appear convincing. But what exactly are we trying to measure?

If the question concerns the repeatability of individual sensors, the repeated measurements are relevant. If we want to understand variation between sensors, however, a hundred readings from the same device do not represent a hundred independent devices. Treating all thousand observations as independent sensor-level replicates could substantially underestimate uncertainty and produce misleading statistical conclusions.

A student familiar with experimental design might immediately question the analysis. Another, despite working with the same data and AI tool, may accept the figure without recognising the problem. Both can produce equally polished reports, yet their understanding of the evidence may be very different.

This creates a challenge for teaching and assessment. AI may reduce differences in the quality of students' immediate outputs while making differences in their underlying understanding less visible. If we evaluate only the finished work, we risk overlooking students who need further support in developing the knowledge required to assess it critically.

I do not expect students entering an interdisciplinary research group to have identical technical backgrounds. They may be highly capable in their own disciplines while having limited experience in others. Effective supervision should recognise these differences and help students develop the foundations needed to make independent judgements. In this example, I would want students to explain what the experimental unit is, which sources of variation matter, and what conclusions the available measurements can support.

Access to AI tools also deserves attention. Students should not be disadvantaged simply because they cannot afford a premium subscription or additional computing resources. Providing appropriate shared resources is one part of the response; ensuring that students have the guidance needed to evaluate AI-generated work is another.

The educational challenge is to make AI a genuine opportunity for learning, rather than a tool that conceals differences in understanding behind equally convincing results.

7. Abdication: allowing the generated account to carry the responsibility

AI can help analyse data, prepare figures and write research papers, but the responsibility for scientific claims remains with the researchers who publish them. This may sound obvious, yet it is surprisingly easy to lose track of important decisions when much of the analysis and writing is automated.

Consider a hypothetical robotic grasping study. A report states that the robot succeeded in 18 out of 20 trials, achieving a 90% success rate. Looking at the original experimental log, however, we discover that 30 attempts were initiated. Ten were excluded because the robot failed to reach the object. The reported 90% may be a valid measure of success once the grasping stage was reached, but the overall task success rate is 60%. Both figures provide useful information, provided we understand what they measure. The problem arises when exclusions are made without justification or when a conditional success rate is presented as overall performance.

Similar issues arise during data processing. Suppose an AI-assisted analysis identifies a sudden drop in gripping force as an outlier and recommends removing it. On closer examination, the synchronised video shows that the object slipped at precisely that moment. Removing the observation would eliminate a genuine failure event from the analysis. The researcher must determine whether an unusual measurement reflects an experimental error or meaningful physical behaviour, rather than accepting an automated recommendation.

Even language editing can alter scientific meaning. There is a substantial difference between saying that changes in contact conditions may be associated with increased error and claiming that they are the primary cause of increased error. The second statement introduces a causal interpretation that may not be supported by the experiment. More confident writing does not make the underlying evidence stronger.

The same responsibility extends to citations and research confidentiality. A reference may exist without supporting the claim attributed to it, while unpublished data and collaborators' manuscripts should not be shared with external AI services without appropriate permission and consideration of data-handling requirements.

I want students to understand that using AI does not reduce their responsibility for the work they produce. They should be able to explain how the data were processed, why particular observations were included or excluded, and what evidence supports their conclusions. When a reviewer questions a result, "AI suggested it" is not a scientific justification.

The researcher must remain accountable for the decisions that turn experimental observations into scientific claims.

What to keep, and what to question

The seven sins are not reasons to discourage students from using AI. These tools can make difficult concepts more accessible, help researchers work across unfamiliar disciplines and remove technical barriers that contribute little to learning. The question is how we use these capabilities without losing the intellectual work that makes education and research meaningful.

The design of AI assistance matters. In a randomised study involving undergraduate physics students, Kestin and colleagues [3] reported greater learning gains with a purpose-built AI tutor than with the particular active-learning classroom condition examined in their experiment. This finding should not be interpreted as evidence that AI tutors are generally superior to teachers. It does, however, suggest that the way assistance is designed and integrated into learning deserves careful attention.

I would like students to approach AI with both curiosity and critical judgement. When learning a new concept, they can use it to explore different explanations, but should still attempt to apply the reasoning independently. When analysing data, they should understand the assumptions behind the methods and be able to explain what the results actually establish. When developing research ideas, they can use AI to explore alternatives, but should resist accepting a proposed direction simply because it sounds convincing or fashionable.

There are practical ways to encourage these habits. Students might make an initial prediction before consulting AI, compare generated explanations with original sources, or ask the tool to identify weaknesses in an argument they have developed themselves. In research, they should be prepared to explain which AI-generated suggestions they accepted or rejected, how they verified consequential outputs, and why their final conclusions are justified.

As educators and supervisors, we should create opportunities for students to practise these decisions, rather than assess only the quality of their finished work. That means recognising careful reasoning, constructive disagreement and the ability to acknowledge uncertainty. It also means providing appropriate guidance, feedback and access to resources so that students with different backgrounds can benefit from AI without becoming dependent on its judgement.

I want students to become confident users of AI, but equally confident in questioning it. They should remain curious about what they do not understand, willing to challenge established explanations and capable of developing ideas of their own.

References and further reading

[1] Bastani, H. et al. (2025). "Generative AI without guardrails can harm learning: Evidence from high school mathematics." Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122

[2] Doshi, A. R. and Hauser, O. P. (2024). "Generative AI enhances individual creativity but reduces the collective diversity of novel content." Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290

[3] Kestin, G. et al. (2025). "AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting." Scientific Reports, 15, 17458. https://doi.org/10.1038/s41598-025-97652-6

Declaration: AI tools were used solely to assist with language editing. All ideas, arguments and intellectual contributions are the author's own.

Illustrations: generated with AI for this post. They are conceptual images and do not depict specific people, apparatus or experimental results.

Join the discussion

What has AI helped you understand—and which suggestions have you had to question? Share your perspective in the comments below. Use the Like button to show your appreciation.

Share on LinkedIn