TechDispatch Review November 2026
In this post, we’ll explore the article TechDispatch, Human Oversight of Automated Decision-Making which was written in 2025.
https://www.edps.europa.eu/system/files/2025-09/25-09-15_techdispatch-human-oversight_en.pdf
While the article was written for Automated decision-making (ADM) and not exclusively for “AI”, many of the concerns remain the same.
I do see that laws like Article 14 in the EU AI Act are written to address many of these concerns.
Source: https://artificialintelligenceact.eu/article/14/
It does call for oversight of foreseeable risks. There are, however, some real time challenges that will take some considerate product design in order to truly minimize the risks.
I found the examples of automation bias alarming. In TechDispatch, the article calls out how the decision-making process became a co-dependent decision rather than human oversight.
Section 3.2 Example: In the 2023 Radiology study, radiologists using AI assistants has a decline in accuracy.
In this case the timeliness of making the correct medical decision does not allow for these types of errors. It can lead to either delayed treatment or unnecessariy treatment being administered based on inaccurate outputs.
Section 3.3 Example: Basque country integrated EPV-R to “evaluate the likelihood of severe intimate partner violence”
”In some instances, the algorithmic recommendations were followed without critical evaluation, potentially reinforcing existing biases and leading to decisions that might not align with the best interests of the individuals involved.”
Without questioning the outputs the judicial system was actually reinforcing the existing biases of the data set used for training and leaving less room for evaluation before judgement.
Both sections also notes that these systems are build with a goal of automation and we assume that there is a timely mechanism for human intervention. These tools have to be mindfully designed, with negative goals taken into account during initial roadmapping process.
The article discusses this in secion 4.1 with the example of social media content moderation. The work of reviewing sensitive, triggering and harmful content is outsourced to workers commonly in “sub-optimal work conditions” with minimal training. Further, the workers receive minimal counseling support for the work done which effects both their ability to perform job and maintain relationships outside of work. When the bottom line is automation and cutting costs, the need for training, counseling services and support mechanisms can be seen as the opposite of automation benefits.
By devaluing the human input in the process of automation, we lose sight of the intended market for the outputs. I believe the course correction is a values shift on human oversight. The TechDispatch report notes that meaningful human oversight can be achieved through:
auditing -regular external auditing to monitor quality and accountability concerns.
sampling - review a subset of the system outputs to monitor performance, detect errors and assess fairness
four-eyes principle - provides redundancy in review that can account for detecting mistake and individual cognitive biases as well as provide transparency of oversight
awareness checkpoints - brief prompts to recalibrate the operator’s situational awareness.
These mechanisms can help build trust by embedding mechanisms of distrust (Braithwaite, 1998).