Code auf Knopfdruck

Code auf Knopfdruck

We have developed a haptic nudge at eye level, right at the edge of the programmer’s screen – the point of decision. Here, we place a multifaceted reminder. On the front of the card are the three Vs – Validate – Verify –Vouch for it – reminding the programmer directly within the active workflow not to blindly trust the AI’s answers, but to question and test them.

Accept the cookie setting to see Panopto content.

Unlock Panopto

What does the topic mean?

Programmers trust the code generated by LLMs like ChatGPT, Gemini, or Claude.
However, this carries risks – resulting in financial damages in the billions, loss of
customer data and trust, as well as security vulnerabilities regarding sensitive data.
The two main reasons are overreliance – the blind trust in AI answers due to time
pressure, convenience, or ignorance – and hallucinations, which are not simply
errors by the AI, but false or fabricated answers presented in a charismatic and
competent manner. Often, sources are made up or existing files are invented. This
interplay generates these massive damages across the business sector, when
human carelessness meets machine failure like Mentos and Coke.

Goal of the nudge

The goal of the nudge is to reduce trust in AI answers and thus prevent
overreliance. We are nudging the human agent holding ultimate accountability –
meaning the hallucinations, unfortunately, remain. However, through our nudge,
programmers are now nudged directly in the situation of deciding between trust or
mistrust toward the AI answers, prompting them to pause and understand the
code, verify it, and later be able to stand behind it with a clear conscience. Thus,
we bring humans back to the driver’s seat of the digital project and significantly
prevent the negative economic consequences of overreliance and hallucinations.

    Needs analysis

    • High error costs due to automation
      Routine processes when writing code lead to a decrease in conscious attention, causing critical edge cases to be overlooked and resulting in expensive rework in quality assurance.
    • Lack of barriers in the workflow
      Common development environments (IDEs) often only offer purely visual standard confirmations during the commit process, which are blindly clicked through due to habituation effects and no longer possess any cognitive relevance.
    • Psychological pressure and time efficiency
      The pressure to meet deadlines tempts developers to shorten code reviews in favor of speed, causing the error probability to increase exponentially.

    Cause analysis

    • Overreliance (Excessive trust in the system)

      Developers tend to
      uncritically transfer control responsibility to automated systems, either
      due to workflow convenience, uncertainty regarding the topic, or time
      pressure. Consequently, AI-generated outputs are neither verified nor
      fully understood.

    • Hallucinations (Confident misinformation by the AI)

      When
      generative AI is utilized within the development process, there is a risk
      that it fabricates non-existent code sources, libraries, or syntax errors,
      presenting these mistakes to the developer in an absolutely plausible
      and confident manner. This leads to massive confusion, increases the
      time required for troubleshooting, and sustainably destroys the
      developers' trust in the generated results.

Target Group

The target group primarily includes professional software developers,
programmers, and IT architects who work in agile project structures and perform
regular code integrations. In addition, the intervention is aimed at managers, team
leads, and project managers in these technical areas in order to create broader
acceptance within the company through their strategic involvement, reach more
people, and anchor quality assurance at an organizational level. The nudge is thus
designed equally for executing developers – from junior to senior – as well as for
the management level.

Added value of the nudge

  • Significant increase in code quality

    Through the haptic barrier at the
    Point of Decision and the subsequent structured validation, critical
    errors and AI hallucinations are intercepted before they reach the
    repository.

  • Enormous resource and cost savings

    Early error reduction relieves
    downstream testing phases (QA) and prevents expensive, time
    consuming correction loops in the deployment process as well as
    fundamental damages during the live phase of the product.

  • Broad behavioral change across all hierarchical levels

    The system
    not only sensitizes developers (from junior to senior) to automated
    blind spots, but also provides managers with an effective tool to
    establish a sustainable, company-wide quality culture.

CONTACT US

Eric Ewald

eric.ewald@stud.hshl.de

Louis Hein

louis.hein@stud.hshl.de

Sebastian Müller

sebastian.mueller2@stud.hshl.de

Olaf Potrzasaj

olaf.potrzasaj@stud.hshl.de