This document summarizes a presentation given by Professor Pekka Abrahamsson on how ChatGPT and AI-assisted coding is profoundly changing software engineering. The presentation covers several key points: - ChatGPT and AI tools like Copilot are beginning to be adopted in software engineering to provide code snippets, answers to technical questions, and assist with debugging, but issues around code ownership, reliability, and security need to be addressed. - Early studies show potential benefits of ChatGPT for tasks like software testing education, code quality improvement, and requirements elicitation, but more research is still needed. - Prompt engineering techniques can help maximize the usefulness of ChatGPT for software engineering tasks. Overall, AI

1. How ChatGPT & AI-assisted
Coding Changes Software
Engineering Profoundly
Professor Pekka Abrahamsson
Tampere University, Finland
K E Y N O T E A D D R E S S
The 38th ACM/SIGAPP Symposium On Applied Computing
March-30th, 2023

2. Pekka Abrahamsson
• Dr. Pekka Abrahamsson works as a full professor of software
engineering at the Tampere Univeristy in Finland. He received his PhD
in Software Engineering in 2002 from the University of Oulu. His
research is in the area of emerging software technologies, empirical
software engineering, and the ethics of artificial intelligence.
• Before his current position, he has served as a full professor at the
University of Jyväskylä (Finland), University of Helsinki (Finland), Free
University of Bolzano (Italy), Norwegian University of Science and
Technology (Norway). He also worked at VTT Technical Research
Centre of Finland as a research professor of software technologies.
• He is widely recognized for his academic achievements. He is a
pioneer in the field of research on agile software engineering methods
and processes. Abrahamsson is the most cited researcher in his field in
Finland. He is the first Professor of Software Engineering to be invited
to the Finnish Academy of Science and Letters.
• He has published broadly in his areas of expertise and received many
awards and recognitions. He was recently ranked in the all-time top 1% of
software engineering scientists globally. Arnetminer named him among the
100 most influential software engineering scientists in the world in
2016. Abrahamsson was awarded the Nokia Foundation Award 2007.
He is the Software Startup Research Network (SSRN) co-founder and
a seasoned expert in leading large research projects.
• His h-index is 62 and he has more than 15600+ citations (March 2023)

3. Shocking news!
• “There is a general agreement that the state of art in practice
[in software industry] is unsatisfactory.
• This state is often described by the term “software crisis”
referring to the poor quality of systems, excessive costs,
schedule and budget overruns.
• It is suggested that the problems lie not in the lack of
methods, techniques or tools.
• We agree and suggest that the fundamental problem is the
limited understanding of system design and its basic
principles.”

4. Shocking news.. 35 ago..
• “There is a general agreement that the state of art in practice [in
software industry] is unsatisfactory.
• This state is often described by the term “software crisis”
referring to the poor quality of systems, excessive costs, schedule
and budget overruns.
• It is suggested that the problems lie not in the lack of methods,
techniques or tools.
• We agree and suggest that the fundamental problem is the limited
understanding of system design and its basic principles.”
Source: Iivari, J. & Koskela, E. (1987): “The PIOCO Model for
Information Systems Design”, MIS Quarterly, 11(03). Pp. 401-419

5. Universal Solution Fallacy
We should have known this?
Malouin, J. L. and M. Landry (1983). "The
mirage of universal methods in systems
design." Journal of Applied Systems
Analysis 10: 47-62.
New method/technology

6. (Ongoing) Misconceptions in the field
• Dependable large systems can only be attained through rigorous
application of the engineering design process
• The key design objective is an architecture that meets
specifications derived from knowable and collectable
requirements
• Individuals of sufficient talent and experience can achieve an
intellectual grasp of the system
• The implementation can be completed before the environment
changes very much
Source: Denning, P.J., Gunderson, C. and Hayes-Roth, R., 2008.
The profession of IT Evolutionary system development.
Communications of the ACM, 51(12), pp.29-31.

7. 7
A State-of-the-art process:
Preparing Ditalini with flageolet & pesto
Unreliable
source
Unfamiliar
terminology
Ambiguous
instructions
Confusing
measures
Incomplete
instructions
Unclear
goal

9. Manipulatibity
Safety
Vulnerability
Volalitility
Robustness
Sustainability Depentability Friendliness Shameability
Pleasurability Substitution of human contact
Normative recognition Data quality
Moral de/re/upskilling Alientation Dignity
Virtuousness Trustability
Benevolence Care concerns Abusability
Responsibility Value sensitivity Malevolence Lethality
Maleficence
Fairness Unpredictability Social sorting
Social solidarity Universal service
Respect for autonomy
Legality
Consent
Access to data
Data collection limitation
Privacy Foreseeability
Predictability
Deceptability Liability
Transparency Righteousness
Blamability
Biasness
Source: Vakkuri, V. and Abrahamsson, P., 2018. The key concepts of ethics
of artificial intelligence. In 2018 IEEE International Conference on
Engineering, Technology and Innovation (ICE/ITMC) (pp. 1-6). IEEE.

10. Summary: What makes software engineering
so hard?
• We are falling short in all the key areas of software engineering
• Requirement gathering and management
• Technical debt
• Integration and interoperability
• Security and privacy
• Scalability and performance
• Testing and quality assurance
• Talent shortage
• We rely too much on human effort in software development. More
than 80% of the code today is still manually entered.

19. 211 companies
were surveyed.
It is a jungle out
there…
For Ethically Aligned AI Development
Source: Vakkuri, V., Kemell, K.K., Jantunen, M., Halme, E. and Abrahamsson, P.,
2021. ECCOLA—A method for implementing ethically aligned AI systems. Journal
of Systems and Software, 182, p.111067.
Download your copy at bit.ly/eccola-method

20. Eary personal experimentation in Jan/2023

22. Code completion
tools
• Microsoft’s Copilot uses Large
Language Model called Codex,
developed by OpenAI, based on
GPT-3
• Trained on Github code
• Works as a developer’s assistant
(pair programmer)
• Focused only on code
• May introduce errors
• 55% increase in productivity (1
study) Source: Pudari, R. and Ernst, N.A., 2023. From Copilot to Pilot:
Towards AI Supported Software Development. arXiv preprint
arXiv:2303.04142.

24. Source: Dall-e generated photos

25. Maybe
ChatGPT (and
language
models) are
just a hype?

26. https://futureoflife.org/open-letter/pause-
giant-ai-experiments/
• Therefore, we call on all AI
labs to immediately pause
for at least 6 months the
training of AI systems more
powerful than GPT-4. This
pause should be public and
verifiable, and include all
key actors. If such a pause
cannot be enacted quickly,
governments should step in
and institute a moratorium.

27. ChatGPT factsheet
• A chatbot, developed by OpenAI company, based in the US, operations funded by
Microsoft by a significant degree
• Built on top of the Large Language Models (LLMs), GPT-3.5, GPT-4
• 100 million+ users, 25M daily
• GPT-3.5 has 170 Billion parameters, GPT-4 has something between 400-1000B (not
confirmed)
• It is now estimated to produce a volume of text every 14 days that is equivalent to all
the printed works of humanity.
• -Source: Dr Thompson, Feb/2023, cited in report by the National Bureau of
Economic Research (Scholes, Bernanke, MIT)

29. GPT-4 promiseware
• GPT-4 accepts both image and text inputs (note! output is in text only today)
• Some Demo’ed Applications:
• GPT-4 can convert your hand-drawn website mockups into actual website code.
• See your refrigerator contents and tell you recipes you can make.
• Read the tax code and calculate your taxes while citing sources.
• GPT-4 outperforms ChatGPT (GPT 3.5) on most academic and professional exams taken by
humans like SAT, GRE, Bar Exams, etc.
• GPT-4 scored in the 90th percentile on the Uniform Bar Exam compared to GPT-3.5, which
scored in the 10th percentile.
• GPT-4 is 82% better than ChatGPT/GPT 3.5 at detecting inappropriate requests and has better
guardrails.
• ChatGPT plugins will be a game-changer for GPT allowing it to talk to external apps like Zapier,
Wolfram, Code interpreters, etc. Open AI may have ushered in a new era of AI app stores.

32. 15 ways to benefit from ChatGPT
Natural Language
Understanding
Multilingual
Conversations
Knowledge Base Creative Writing Problem Solving
Simulating
Conversations
Personalized
Recommendations
Summarization
and Simplification
Debates and
Perspectives
Code and
Technical Help
Role-playing and
Gaming
Learning and
Education
Emotional Support
Language
Translation
Grammar and
Writing Assistance

33. How ChatGPT is argued to help software
engineers?
1.Providing answers to technical questions: Software engineers often encounter complex technical
problems that require research and analysis. ChatGPT can provide quick and accurate answers to
these questions, drawing on a vast repository of knowledge.
2.Generating code snippets: ChatGPT can also generate code snippets for specific tasks, which can
save software engineers time and effort. This can be particularly useful for common tasks or for
code that follows a specific pattern.
3.Assisting with debugging: ChatGPT can help software engineers identify and troubleshoot issues
in their code by analyzing error messages and providing suggestions for fixes.
4.Offering insights on emerging technologies: ChatGPT can keep software engineers up-to-date
with the latest trends and advancements in their field, such as new programming languages,
frameworks, or tools.
5.Supporting collaboration: ChatGPT can help facilitate collaboration among software engineers by
providing a platform for real-time communication and sharing of ideas and resources.

34. Known issues / challenges
• There are several problems with the use of ChatGPT, Copilot and
others, which need to be solved before wider adoption:
• Code ownership, IPR issues
• Limited applicability scope (limited due to training data)
• False instructions, advice, information
• Code defects
• Known and unknown security threats
• Security and privacy concerns
• Working in a client development environment
• Difficulty in integrating with an existing workflow and tools
• Costs of large language models can be very high

35. ChatGPT’s own
advice with IPR
issues

36. 36
Common Use Cases
AI-Assisted learning /
Project onboarding /
Training / Personal
assistant
Use Case 1
AI-Assisted Software
Engineering /
Development
Use Case 2
AI-Assisted Decision
Making based on your
own data
Use Case 3

37. What do the scholars say now?
• ~1000 papers on Large Language Models in Arxiv (as of March-28th)
• 52 papers on LLMs and software engineering
• General themes covered: Program Synthesis, AI Evaluation, Bug Detection, Error
Handling, Learning Materials Generation, Code Analysis, Code Completion Systems,
Reverse Engineering, Spreadsheet Models and Code Poisoning
• 170 articles on ChatGPT or employed ChatGPT in Arxiv
• 90 articles with ChatGPT on title
• Only three studies related to Software Engineering
• ChatGPT and Software Testing Education: Promises & Perils (experiment)
• Towards Human-Bot Collaborative Software Architecting with ChatGPT (case study)
• ChatGPT Prompt Patterns for Improving Code Quality, Refactoring, Requirements
Elicitation, and Software Design (experience-based)

38. Example Prompt engineering patterns for SW development
Source: White, J., Hays, S., Fu, Q., Spencer-Smith, J. and Schmidt, D.C., 2023. ChatGPT Prompt
Patterns for Improving Code Quality, Refactoring, Requirements Elicitation, and Software Design.
arXiv preprint arXiv:2303.07839.

39. Example Prompt engineering patterns for SW development
Source: White, J., Hays, S., Fu, Q., Spencer-Smith, J. and Schmidt, D.C., 2023. ChatGPT Prompt Patterns for Improving Code Quality, Refactoring, Requirements Elicitation,
and Software Design. arXiv preprint arXiv:2303.07839.

40. Studied themes varied greatly
• Virtual Reality and Metaverse
• Translation Evaluation
• Machine Translation
• Ethics and Regulation
• Academic Publishing
• Plagiarism Detection
• AI Generated Content
• Bug Fixing
• Bioinformatics
• Sentiment Analysis
• Medical Advice
• Construction Project Scheduling
• Software Testing Education
• Large Language Model Failures
• Statistical Process Control
• Designer AI
• Ordered Importance Communications
• Learning Gain Comparison
• Zero-Shot Information Extraction
• Causal-Discovery Performance
• AI Ethics

41. Some empirical findings
• ChatGPT was able to respond correctly to 56% of Software Testing exam questions, Jalil et al,
2023
• ChatGPT narrowely passed a computer science exam (24/40, student average 24), Bordt and
von Luxburg, 2023
• ChatGPT resembles closely human patterns in language use, Cai et al, 2023 (10/12
experiments passed)
• ChatGPT's ranking preferences are quite consistent with human, Ji et al, 2023 (can be used to
categorize data, zero-shot ranking capability good)
• ChatGPT beats Grammarly in fixing grammatical errors, Wu et al, 2023
• ChatGPT’s zero-shot Text-to-SQL capabilities are impressively good, Liu et al., 2023
• ChatGPT is an excellent Keyphrase generator, Song et al, 2023
• ChatGPT lacks moral authority and is not consistent in its advice, Krügel et al, 2023
• ChatGPT is already at commercial product level in language translation, Jiao et al, 2023
• ChatGPT is 20x less costly than M-Turk for text annotation tasks and more accurate, Gilardi et
al., 2023

42. Conducting Systematic Literature Reviews with ChatGPT: A
Proposal
Source: Waseem, M., Ahmad, A., Liang, P., Fehmideh, M., Abrahamsson, P.
and Mikkonen, T., Conducting Systematic Literature Reviews with ChatGPT,
2023, Researchgate

43. Final thought, a new must-have skill for you
all, the art of Prompt Engineering

44. Key messages
• Despite of advances, software engineering continues to be in crisis
• Adoption of AI-assisted tools is still in its infancy
• Introduction of LLMs may be a game changer in the field of SE but also in other
fields as well.
• ChatGPT offered the missing user interface for the use of AI in various contexts.
While scientific studies are still coming, early results indicate positive influences
across many sectors.
• It may hot air as well
• Assistant that delivers 50% false results and provides a different answer to every question,
would get fired in real life
• Ethics issues are real, training material is biased
• Yet I believe that we should explore the new AI tools such as ChatGPT will full
force
• The question remains, how ChatGPT will help you research?

45. Reach me at
pekka.abrahamsson@tuni.fi