03 · What You Need to Know
Why Competing Explanations Matter in Research
The first plausible explanation is rarely the only possible one
Researchers routinely move from observations toward explanations. Students using a tool perform differently. An intervention is associated with better outcomes. One group reports higher satisfaction. A policy change is followed by a behavioral change.
The difficulty is that the observed pattern usually does not contain its own explanation.
If two variables are associated, several causal structures may be compatible with that association. X may influence Y. Y may influence X. Another variable may influence both. Selection processes may create the relationship. Measurement artifacts may produce it. The relationship may also differ across conditions that have not yet been considered.
Generative AI can be useful because producing alternatives is precisely a generative task. Research has already examined LLMs for scientific hypothesis generation, including systems that generate hypotheses from data and open-domain scientific material. These studies suggest that LLMs can produce potentially useful candidate hypotheses, while also showing why generated hypotheses require independent evaluation.
An alternative explanation is not necessarily a competing theory
Researchers sometimes use “alternative explanation” loosely, but alternatives can occur at several levels.
| Type of alternative |
Question it raises |
Example |
| Competing mechanism |
Could another process produce the same outcome? |
AI use affects writing through feedback frequency rather than reduced cognitive effort |
| Reverse direction |
Could the presumed outcome partly cause the presumed predictor? |
Students who already struggle with writing may use AI more often |
| Confounding |
Could another factor influence both variables? |
Prior achievement influences both AI use and subsequent performance |
| Selection |
Could who enters the observed groups explain the pattern? |
Students choosing AI assistance differ systematically from non-users |
| Measurement |
Could the pattern reflect how variables were measured? |
Self-reported AI use poorly represents actual use |
| Contextual condition |
Could the relationship change across settings or circumstances? |
AI effects differ between formative and high-stakes assessment |
| Theoretical interpretation |
Could another framework explain the same behavior differently? |
AI reliance could be interpreted as dependence or as strategic resource use |
AI can help populate these categories. The researcher still needs to determine which alternatives are scientifically plausible and consequential.
Alternative explanations can improve research design
Generating rival explanations is not merely an intellectual exercise. It can change what evidence you collect.
If prior achievement could explain the apparent relationship between AI use and performance, you may need baseline achievement data. If reverse causality is plausible, temporal ordering becomes important. If instructional quality could influence both variables, the sampling or modeling strategy may need to account for it.
A competing explanation therefore becomes useful when it forces you to ask what observation would distinguish one account from another.
Interesting alternative
Another plausible story about why the observed phenomenon might occur.
Research-useful alternative
An alternative that changes what evidence, comparison, measurement, design, or prediction would help distinguish competing accounts.
AI can deliberately reverse your presumed causal story
Researchers often write questions with a causal direction already embedded in them.
“How does AI dependence affect critical thinking?” presumes that dependence is the explanatory factor and critical thinking is the outcome.
You can ask AI to reverse the relationship: could people with lower confidence in their independent reasoning become more dependent on AI? Could both processes operate simultaneously? Could another characteristic influence both?
This does not mean every reversed explanation is equally credible. It prevents the grammar of your research question from quietly becoming causal evidence.
AI can suggest confounders you have not considered
Suppose students who use generative AI heavily also earn lower grades. Before attributing the difference to AI, consider what distinguishes heavy users from light users.
Prior academic performance, workload, language proficiency, course difficulty, motivation, digital skills, access to tutoring, assessment type, instructor policy, or other factors could be relevant depending on the context.
Generative AI can quickly produce a long list of possibilities. This is helpful during exploratory planning, but a long list of imaginable confounders is not equivalent to a causal model.
Potential variables should be justified using subject-matter knowledge, theory, prior evidence, and the causal assumptions appropriate to the design. Blindly controlling for every variable AI suggests can itself introduce bias.
AI can propose alternative theoretical interpretations
The same observation can mean different things under different theoretical frameworks.
Consider students repeatedly consulting generative AI before completing a task. One interpretation might frame this as technological dependence. A self-regulated learning perspective might ask whether students are strategically seeking external assistance. A distributed cognition perspective might question whether the boundary between “independent” and tool-assisted cognition has been defined appropriately.
These perspectives do not simply rename the behavior. They can imply different mechanisms, measurements, and predictions.
AI can help researchers encounter theoretical vocabularies outside their usual intellectual neighborhood. Yet theoretical compatibility must be checked carefully. A model can combine theories because their language appears related even when their assumptions do not fit together.
Alternative explanations can reveal that the original problem was framed too narrowly
Suppose a researcher begins with the claim that faculty members avoid generative AI because they lack AI literacy.
Alternative explanations might include pedagogical disagreement, privacy concerns, assessment incompatibility, disciplinary norms, institutional policy uncertainty, workload, or deliberate professional judgment.
If several alternatives are plausible, the original problem statement may have prematurely converted one explanation into the problem itself.
This is why generating alternatives can complement the earlier task of identifying the underlying research problem. Sometimes the most useful outcome is discovering that you do not yet know what the problem actually is.
AI-generated alternatives can themselves be wrong
Generative models can produce scientifically plausible propositions without reliable grounding. Research on LLM hypothesis generation has explicitly identified hallucination and truthfulness as important challenges. One 2025 IJCAI study evaluating scientific hypothesis generation reported that LLMs struggled to generate truthful hypotheses reliably, illustrating why plausibility should not be confused with factual support.
Watch Out
An alternative explanation does not become scientifically credible because AI can articulate a mechanism for it. The mechanism may contradict established evidence, rely on nonexistent relationships, misuse a theory, or simply be invented.
This limitation does not make hypothesis generation useless. Hypotheses are propositions to test. It does mean that the researcher must distinguish productive speculation from evidence-backed explanation.
Ask for discriminating predictions, not only alternative stories
A particularly useful next step is to ask what evidence would differ if Explanation A were true rather than Explanation B.
Suppose one explanation predicts that frequent AI use reduces independent problem solving, while another predicts that students already struggling with independent problem solving seek AI more often.
These accounts imply different temporal patterns. Measuring problem-solving ability before substantial AI use, tracking subsequent use, and observing later performance may provide evidence unavailable from a one-time correlation.
The stronger question is therefore not merely “What else could explain this?” but “What observation would help distinguish these explanations?”
Use AI to increase explanatory competition, not to manufacture certainty
Scientific reasoning benefits when a favored explanation has credible rivals. If every observation is interpreted only through the researcher's preferred theory, evidence can become strangely cooperative.
Generative AI can make rival-generation inexpensive. You can ask for the strongest explanation consistent with your theory, then the strongest explanation inconsistent with it. You can ask how a researcher from another discipline might interpret the same observation. You can ask which alternative would be most damaging to your intended conclusion if true.
This moves naturally toward using AI to challenge assumptions embedded in the research idea.
04 · A Practical Example
One Pattern, Five Possible Explanations
Hypothetical Example
Frequent AI users perform worse on an independent writing task
A researcher analyzes student data and observes that students reporting frequent generative AI use obtain lower scores on an unaided writing assessment. The researcher's first interpretation is that regular AI use may be weakening students' independent writing ability.
Before developing the study around that explanation, the researcher asks AI to generate substantively different accounts of the same association.
Explanation 1: AI-use effect
Frequent reliance on generated text may reduce opportunities for independent writing practice.
Explanation 2: Reverse direction
Students who already have difficulty writing independently may be more likely to seek AI assistance.
Explanation 3: Third-variable explanation
Students with heavier workloads may use AI more frequently and also have less time to prepare for the assessment.
Explanation 4: Measurement problem
“Frequency of AI use” may combine fundamentally different practices, such as requesting explanations, checking grammar, brainstorming, and generating complete answers.
Explanation 5: Contextual explanation
The relationship may occur mainly in courses where students receive little structured writing support, rather than being a general consequence of AI use.
The researcher now has a different problem. Instead of trying to demonstrate one preferred explanation, the study can be designed to distinguish among plausible alternatives. That is a considerably stronger position from which to investigate the phenomenon.
