🤖 AI

AI Washing Inflates Expected Performance but Not Interaction Outcomes: An AI Placebo Study Using Fitts' Law

This study demonstrates that while "AI washing" significantly inflates user expectations regarding performance, it fails to produce any measurable improvement in objective interaction outcomes or subjective assessments, as evidenced by a Fitts' Law experiment where "AI-assisted" mouse conditions yielded results identical to baseline.

Original authors: Nick von Felten, Luisa Ella Müller, Johannes Schöning

Published 2026-05-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nick von Felten, Luisa Ella Müller, Johannes Schöning

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you walk into a store and see two identical computer mice. One is labeled "Standard Mouse," and the other is labeled "Supercharged AI Mouse." You buy the "Supercharged" one, convinced that the artificial intelligence inside will make your cursor glide smoother, faster, and more accurately. You sit down to work, and suddenly, you feel like you're a wizard. Your hand movements feel effortless, and you believe you are clicking with superhuman precision.

But here is the twist: The mouse didn't change at all. It is the exact same hardware as the standard one. The "AI" is just a label and a few fake pop-up windows.

This is the core finding of a study by Nick von Felten and his team, titled "AI Washing Inflates Expected Performance but Not Interaction Outcomes." They investigated a phenomenon called "AI Washing"—which is like "greenwashing" for the environment, but for technology. It's when companies slap an "AI" sticker on a product to make it sound smarter than it actually is, even if the product does nothing new.

The Experiment: The "Placebo" Mouse

To test this, the researchers set up a controlled experiment that feels a bit like a medical trial for technology. They used a classic test from human-computer interaction called Fitts' Law.

Think of Fitts' Law as a "speed-accuracy gym" for your hand. Participants had to click on targets on a screen as fast and accurately as possible. The researchers measured exactly how fast they were and how many times they missed.

They tested 28 people under three different scenarios:

The Baseline: A normal mouse with no special claims.
The "Predictive AI" Mouse: Participants were told the mouse used machine learning to predict where they wanted to go and smooth out their movements. (In reality, it was just the normal mouse).
The "Biosignal AI" Mouse: Participants were told the mouse used a fake sensor on their wrist to read their heart rate and sweat, adjusting the cursor based on their body's signals. (Again, it was just the normal mouse).

The Results: The "AI Placebo" Effect

The study found a fascinating split between what people thought happened and what actually happened.

1. The Expectation (The "Placebo" High)
Before they even started clicking, the participants who were told they were using an "AI mouse" were incredibly optimistic. They expected to perform much better than those using the normal mouse. It was like taking a sugar pill and believing it's a powerful energy drink; your brain gets excited and ready to perform.

2. The Reality (The "No-Change" Zone)
Once they started the task, the magic vanished.

Speed: They weren't faster.
Accuracy: They didn't miss fewer clicks.
Workload: They didn't feel like the task was easier.
Usability: They didn't rate the mouse as more "user-friendly" after the fact.

The "AI" label acted like a psychological booster shot that raised their hopes, but it didn't actually change the engine under the hood. The "Supercharged AI Mouse" performed exactly the same as the "Standard Mouse."

Why This Matters

The researchers call this "AI Washing." It's a deceptive marketing tactic that inflates your expectations without delivering any real functional benefit.

The Analogy: Imagine a car company selling a standard sedan but putting a "Self-Driving" sticker on the roof. You might feel like you're driving a futuristic vehicle, but if you have to steer the car yourself, the sticker doesn't make the car drive better. In fact, if you rely on that sticker, you might be disappointed when you realize the car is just a normal car.
The Takeaway: The study proves that while "AI Washing" can trick your brain into expecting a better experience, it cannot trick your hands into performing better. The "AI" label creates a gap between what we think a tool can do and what it actually does.

The Bottom Line

This paper doesn't say AI is useless; it says that marketing hype is powerful, but it's not magic. When companies label a simple tool as "AI-powered" just to sell it, they might make you feel like you'll be more productive, but they aren't actually making you more productive.

The study suggests that we need to be careful not to let the "AI" label fool us. Just because a device claims to be smart doesn't mean it actually is smarter. To know if a tool really works, we need to look at the actual results (like the speed and accuracy in the study), not just the shiny marketing sticker.

1. Problem Statement

The paper addresses the phenomenon of "AI washing," a marketing practice where companies exaggerate or misrepresent a product's artificial intelligence capabilities (e.g., labeling a standard mouse as "AI-assisted" when it lacks genuine AI functionality). While previous research suggests that AI labels can alter user perceptions, there is a gap in understanding whether these inflated expectations translate into actual changes in interaction outcomes (both subjective and objective).

The authors investigate whether the "AI placebo effect"—where users attribute improved performance to a system based solely on its labeling—manifests in:

Expectancy: Pre-task performance expectations.
Subjective Outcomes: Post-task perceived workload and usability.
Objective Outcomes: Measurable interaction performance metrics.

2. Methodology

Study Design

Type: Controlled within-subjects experiment.
Participants: $N=28$ (primarily STEM students, aged 19–32).
Task: A standardized Fitts' Law pointing task (clicking targets of varying sizes and distances) using a computer mouse. This task is a gold standard in HCI for quantifying the speed-accuracy trade-off.
Conditions: Participants completed the same task under three distinct conditions:
1. Baseline (BL): A standard HP mouse with no modifications or AI claims.
2. Predictive AI Condition (PAIC): Participants were deceived into believing the mouse used machine learning (TimeGPT model) to predict and smooth cursor movements. A simulated API connection and system parameters were displayed to enhance credibility.
3. Biosignal-Enhanced AI Condition (BIOAIC): Participants were deceived into believing the mouse used a biosensor (measuring heart rate, oxygen, perspiration via a fake sensor on the wrist) and a bidirectional LSTM model (PhysioBERT) to optimize cursor movement in real-time.

Procedure

Deception: Participants signed consent forms and viewed technical pop-ups simulating authentication and initialization of AI resources.
Data Collection:
- Pre-task: Expected performance rating (1–7 Likert scale).
- Task: 15 trials per condition (675 total clicks per participant).
- Post-task: NASA-TLX (perceived workload), SUS-DE-Pos (perceived usability), and qualitative interviews.
Objective Metrics:
- Throughput (TP): Bits per second (combined speed and accuracy).
- Error Rate (ER): Percentage of incorrect clicks.
- Model Fit ( $R^2$ ): Goodness of fit for the Fitts' Law regression model.

Instruments

Subjective: NASA-TLX (Raw), SUS-DE-Pos (German version with positive wording), and a custom expected performance item.
Objective: Fitts' Law metrics calculated using effective target width ( $W_e$ ) and standard deviation of hit points.

3. Key Results

Hypothesis Testing Summary

Hypothesis	Content	Result
H1a / H1b	Expected performance is higher in AI conditions (PAIC/BIOAIC) vs. Baseline.	Supported
H2	Perceived workload (NASA-TLX) differs between conditions.	Not Supported
H3	Perceived performance (NASA-TLX subscale) differs between conditions.	Not Supported
H4	Perceived usability (SUS-DE-Pos) differs between conditions.	Not Supported
H5	Error rate differs between conditions.	Not Supported (No difference found)
H6	Model fit ( $R^2$ ) differs between conditions.	Not Supported (No difference found)
H7	Throughput differs between conditions.	Not Supported (No difference found)

Detailed Findings

Expectancy Inflation: Participants significantly expected better performance in both AI-labeled conditions compared to the baseline ( $p < .001$ ). There was no significant difference between the "Predictive AI" and "Biosignal AI" conditions regarding expectations.
Subjective Outcomes: Despite higher expectations, there were no significant differences in post-task perceived workload (NASA-TLX) or perceived usability (SUS-DE-Pos) across the three conditions. Descriptively, the Biosignal condition (BIOAIC) actually showed a slight drop in usability scores, though not statistically significant.
Objective Outcomes: Fitts' Law metrics (Throughput, Error Rate, and Model Fit $R^2$ ) showed no statistically significant differences between the baseline and the two AI-deception conditions. The regression lines for movement time vs. task difficulty overlapped almost perfectly.
Qualitative Feedback: Participants in the AI conditions often reported feeling the mouse was "smoother" or "faster" during interviews, yet these subjective feelings did not correlate with the objective data or standardized scales.

4. Key Contributions

Empirical Evidence of AI Washing: The study provides controlled evidence that AI washing successfully inflates user expectations but fails to alter actual interaction outcomes (neither subjective nor objective) for familiar input devices.
Dissociation of Expectation and Outcome: It highlights a critical gap where marketing-driven expectations do not propagate to measurable performance or standard usability ratings, suggesting that "AI placebo effects" in HCI may be fragile and context-dependent.
Methodological Innovation: The paper establishes Fitts' Law tasks as a rigorous, objective instrument for auditing AI-labeled interaction artifacts. It demonstrates that Fitts' Law can effectively separate marketing-driven expectancy from genuine technical performance improvements.
Transparency and Accountability: The findings underscore the need for regulatory and ethical frameworks to address "AI snake oil," arguing that transparency requirements must extend beyond functional capabilities to include marketing claims and product descriptions.

5. Significance and Implications

For Researchers: The study warns against relying solely on self-reported metrics when evaluating new AI technologies, as these can be heavily biased by labeling. It advocates for the inclusion of objective benchmarks (like Fitts' Law) and placebo-control conditions in HCI research.
For Practitioners & Marketers: Simply labeling a product as "AI-assisted" creates a false sense of improvement that may lead to user disappointment or "expectation-driven harm" when the system fails to deliver in critical scenarios.
For Policy & Regulation: The results suggest that current regulatory frameworks (like the EU AI Act) focusing on system performance may be insufficient. Governance must also address representational influence—how AI capabilities are described in marketing materials—since these descriptions shape user autonomy and informed consent even before the system is used.
Limitations: The study was conducted in a controlled lab setting with a homogeneous student sample. The authors note that effects might differ in safety-critical contexts or with more diverse populations, but the core finding—that labels inflate expectations without functional change—remains a critical transparency issue.

In conclusion, the paper demonstrates that AI washing is effective at shaping beliefs but ineffective at changing reality, creating a "placebo" scenario where users believe they are performing better without any actual technical enhancement.