AI is everywhere nowadays. Its potential to support our work is exciting, and its results can range from impressive to hilarious, depending on the context. Here is a little challenge for you. Check out the video of this gymnast and then this other one. Can you tell apart which of these gymnasts is a real human and which of them is generated by AI?
In all transparency, it should not be much of a challenge (at least if you are human) because of the obvious perceptual cues (and because the text in the links clearly tells you the answer). But just in case, for the friendly AIs reading this post, the first video portrays a routine by American gymnast Simone Biles (human), and the second one is a video of an AI-generated gymnast that is hilarious enough to be featured in various media.
On a more serious note, one way AI could support real and impactful work is by assisting healthcare professionals in tasks such as detecting whether cancer is present in screening mammograms. Not surprisingly, the rapid developments in AI have prompted policymakers to work on standards for its use, such as striving to increase transparency.
And to be honest, it sounds compelling. But what if, empirically, higher transparency could lead to more negative outcomes? Melina Kunar, Giovanni Montana, and Derrick Watson developed a thought-provoking experiment on the matter.
It is reported in their paper “Increasing transparency of computer-aided detection impairs decision-making in visual search,” recently published in Psychonomic Bulletin & Review, a Psychonomic Society journal.
The experiment
The research team recruited more than 600 participants whose task was to judge whether cancer was present or not in a set of mammographies. To do so, participants received a short training during the introduction to the experiment, which used specially designed stimuli suitable for individuals without medical training to identify cancer.
In addition, participants could have the help of a Computer Aided Detection (CAD) system, which would sometimes mark areas of the mammography that could indicate cancer presence. In such a setting, there are 5 types of responses. If cancer was present in the mammography (see panels below), the CAD system could
- Mark the correct area
- Mark an incorrect area
- Not mark any area
In contrast, in mammographies without signs of cancer (see panels below), the CAD system could:
- Incorrectly mark an area as cancerous
- Not mark any area
The authors manipulated how accurate the CAD system was and whether there was transparency in its accuracy level. The CAD system could be accurate in 33%, 66%, or 83% of the trials, and participants could either be told what the accuracy level was (transparent condition), or not be given any information about it (non-transparent condition). This resulted in six conditions to which different groups of participants were randomly assigned. For completeness, some participants did not have access to a CAD system, which served as a control condition.
Detecting signals of cancer
The matrix below shows four possible outcomes depending on whether the cancer was present or absent in the mammograph and whether the participants responded that cancer was present or not.
| Mammography | |||
| Cancer present | Cancer absent | ||
| Participant’s response | Cancer present | Hit | False Alarm |
| Cancer absent | Miss | Correct rejection | |
If there was cancer in the mammography and the participant responded that there was cancer, the response would be correct (a “hit” in the table above), but if the participant responded that there was no cancer, they would have “missed” it.
In contrast, if there was no cancer in the mammography and the participant responded that there was no cancer, the response would be a “correct rejection,” but if the participant responded that there was cancer, it would be a “false alarm.”
This structure is extremely convenient from a methodological perspective, as it allows us to model the situation using Signal Detection Theory, an approach applied to a wide variety of situations, some of which we have covered before.
The categories described above allow computing indexes of sensitivity (d’ or “d prime”) and the degree of response bias (c or “criterion”) that participants display.
Transparency is better … Or is it?
As shown in the top left panel of the figure below, the research team found that participants with access to more accurate CAD systems committed fewer misses but found no effects of the transparency manipulation. In contrast, the rate of false alarms (top right panel) was higher in the transparent condition than in the non-transparent one, at least for the CAD systems, with 33% and 66% accuracy rates.
The middle left panel in the figure above shows the more accurate CAD systems increased participants’ sensitivity, as evidenced by higher d’ values (i.e., better ability to detect cancer from non-cancer). However, the transparent conditions led to lower d’ values, showing reduced sensitivity (i.e., reduced ability to detect cancer from non-cancer mammography). Lastly, as shown in the middle right panel in the figure above, lower levels of c were observed as the accuracy of the CAD systems increased, revealing that participants were more biased in responding that cancer was present. In addition, the transparent condition led to a bias to respond that cancer was present.
Could non-transparency be a better option?
The results of the study are thought-provoking and certainly call for a closer empirical analysis to determine whether, in terms of outcomes, transparency could indeed be the best choice or not. According to the authors:
“Along with recent developments of Artificial Intelligence (AI) in healthcare, it is also essential to investigate how humans interact with AI to make decisions. This research shows that giving people information about the accuracy of AI leads to impaired performance in a cancer detection task. In this case, increased transparency of the AI leads to over-reliance on the technology, which can impact clinical decision making”.
I usually like to end my posts lightheartedly, but given the importance of the topic, this time, I wouldn’t want to miss the opportunity to end it with an invitation to get screened for breast cancer.
Psychonomic Society’s featured article
Kunar, M.A., Montana, G. & Watson, D.G. (2024). Increasing transparency of computer-aided detection impairs decision-making in visual search. Psychonomic Bulletin and Review. https://doi.org/10.3758/s13423-024-02601-5
