This presentation is a part of Aalto University’s webinar series on

Research Data Management

& Open Science
Autumn 2023

All slides in this presentation are licensed with CC-BY and can be reused with attribution to doi:10.5281/zenodo.8413164, unless explicitly mentioned otherwise.
AI & Research work

Dr. Enrico Glerean, Staff Scientist at SCI and Data Agent

Dr. Pedro Silva, Postdoc at CHEM and Data Agent
10.10.2023
Outline: AI in research work
1. From AI to generative AI Learning outcomes
2. Prompt engineering for - How to use generative AI
research tools in your research work
3. AI in the academic workflow - Evaluating the outputs and
4. Ethical and legal understanding limitations
considerations - Awareness of the larger
5. Other tools and advanced picture of AI tools today
topics
 DISCLAIMER
THIS IS NOT A COMPUTER SCIENCE COURSE ON GENERATIVE
ARTIFICIAL INTELLIGENCE. THE MATERIALS ARE TAILORED FOR
RESEARCHERS WHO HAVE NEVER USED THESE TECHNOLOGIES

SINCE WE ONLY HAVE 1 HOUR, ENRICO AND PEDRO WILL FOCUS

ON INTUITIVE EXPLANATIONS AND PRACTICAL HINTS ON HOW TO
USE THE MOST POPULAR TOOLS AND WHAT TO BE AWARE OF
WHEN USING THEM

SHARE YOUR EXPERIENCE AND HINTS WITH OTHERS SO THAT

WE CAN BUILD AN OPEN GUIDE FOR RESEARCHERS
The Research Process aided by AI
with Responsible Conduct of Research at its core and some examples

RCR
Ethics & Law

Original image from https://rdmkit.elixir-europe.org/ “The data lifecycle”

1. What is (generative) AI?
What is Artificial Intelligence?
Multiple definitions since 1956, let’s pick one from law:
“‘Artificial Intelligence system’ (AI system) means software that is developed
with one or more of the techniques and approaches listed in Annex I (e.g.
machine learning and statistics) and can, for a given set of human-defined
objectives, generate outputs such as content, predictions, recommendations, or
decisions influencing the environments they interact with”
From “Proposal for a REGULATION OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL LAYING
DOWN HARMONISED RULES ON ARTIFICIAL INTELLIGENCE (ARTIFICIAL INTELLIGENCE ACT) AND
AMENDING CERTAIN UNION LEGISLATIVE ACTS”

Other AI legal definitions collected by IAPP.

Learning cats and dogs: Machine learning
1) Learning
stage

2) Testing
stage

From http://www-labs.iro.umontreal.ca/~liubang/files/DogCat_report.pdf
Generative AI: an intuitive explanation
- The key task in machine learning is trying to predict
- The classifier from the previous example is able to predict
labels from new images: given a new picture of a dog, it can tell
if there’s a cat or a dog
- There are other types in AI called “generative”: the program
learns patterns from the training set and is able to generate a
picture, or text, or sound following the rules of what it has learned

We are not even scratching the surface here, if you are interested to know more, ask in the chat
and we can provide references. See this link for a great collection of references
Focus on text generation:
large language models (LLMs)
Why focus on language?
• Large amount of data for training these foundational models
• Language is possibly the most versatile modality as it allows
to describe (almost) anything: concepts, facts, relationships,
processes, software code, …
• Language (as text via a keyboard) is an efficient way to
interact with a computer program via instructions
Guess the letter
It is an English word starting with
the letter 'q': which letter comes
after that?
Guess the next letter
If I am thinking an English word that
starts with q and I ask you to guess
the next letter, what is your most
likely guess?

Image from What is ChatGPT doing, and

why does it work? by Stephen Wolfram
How often a letter from the columns is
followed by a letter on the rows? Bright
yellow = never, dark red = very often
Guess the next word, sentence,
paragraph, chapter, book…
Large language models estimate the probability between words in the context
of sentences, paragraphs, chapters, books… Given a prompt sentence, the
language model can be asked to continue it based on the most likely expected
following words, with a given randomness (temperature)
E.g. the sentence “The best thing about AI is its ability to” has these as the 5
most likely following words: “learn”(4.5%), “predict” (3.5%), “make”, (3.2%),
“understand” (3.1%), “do” (2.9%)

From What is ChatGPT doing, and why does it work? by Stephen Wolfram
Guess the next word, sentence,
paragraph, chapter, book… (II)
{The best thing about AI is its ability to, At each generated word, we re-
The best thing about AI is its ability to create, ask the model to pick another
The best thing about AI is its ability to create worlds, one from the top N words that
The best thing about AI is its ability to create worlds that, would follow from all the
The best thing about AI is its ability to create worlds that are, previous words seen so far.
The best thing about AI is its ability to create worlds that are both,
The best thing about AI is its ability to create worlds that are both exciting,}

The best thing about AI is its ability to learn. I’ve always liked the Different runs, produce
The best thing about AI is its ability to really come into your world and just
different outputs (goodbye
The best thing about AI is its ability to examine human behavior and the way it
reproducibility!)
The best thing about AI is its ability to do a great job of teaching us
The best thing about AI is its ability to create real tasks, buy you can From What is ChatGPT doing, and why does it work?
by Stephen Wolfram
History of LLMs
Intuitive explanations:
- Encoder models: they
“understand” (= map it to a
multidimensional space of
words) the prompt
- Encoder-decoder models:
they “understand” the
prompt and produce a new
response
- Decoder models: they are
usually autoregressive and
only generate based on what
has been given so far.

Yang, J., Jin, H., Tang, R., Han,

X., Feng, Q., Jiang, H., ... & Hu,
X. (2023). Harnessing the power
of llms in practice: A survey on
chatgpt and beyond. arXiv
preprint arXiv:2304.13712.
How to use them?
Platforms currently available
Historically the way to query generative models was to ask “continue the
sentence”. Later OpenAI made the chat interface successful, but they were
not the first (ref)
Most current platform require you to sign-up to a service since the actual
model and its inference software are not running on your local computer.
Current platforms and technologies: OpenAI with free ChatGPT 3.5 and paid
ChatGPT 4 (also Microsoft Bing search engine), Anthropic Claude 2
(requires VPN), Google Bard, Huggingface chat (various open-source
models available)
If privacy and protection of your data is important, open-source models can
also be run locally.
PS: did you hear about AaltoGPT?
2. Prompt engineering

“Prompt engineering is the process of

constructing queries or inputs for AI
language models so as to elicit the most
precise, coherent, and pertinent responses.”
 The art of asking The answer you
were looking for

Do you
have time?

The answer to the

question you asked

Is Herlinda asking:
1. to meet with Tiyna?
2. if Tiyna is available for leisure after work?
3. what is the current time?
Structuring the prompt
1. Start with a concise question:
Direct: What is the capital of France?
Instructional: Translate the following text to French: “(…)”
Exploratory: Discuss the implications of climate change on beetles.
Contextual: Given the economic downturn in 2008, what (…)
Comparative: Compare the economic systems of Finland and Sweden.
Scenario-based: Imagine you are a researcher, how would you defend funding in fundamental research?
…

2. Specify tone and style: ex. write in an academic tone, explain like a 5th grader, explain to
a 5th grader, translate to Finnish, …
3. Add examples as input: this can help the AI towards the answer
4. Iterative refinement: after getting the first answer, you can engage in an iterative dialogue
CLEAR Principles
Explain
1. Concise: brevity and clarity in “Can you provide me with a detailed explanation of the
process of photosynthesis and its significance?”
prompts
“(1) List the steps to write a research paper, (2) beginning
2. Logical: structured and coherent with selecting a topic and (3) ending with proofreading the
final draft.”
prompts
Instead of “What are some renewable energy sources?”,
3. Explicit: clear output specifications ask for “Identify five renewable energy sources and
explain how each works.”
4. Adaptive: flexibility and
If asking, “What are some ways to conserve water?” leads
customization in prompts to generic responses, try a more targeted and adaptive
prompt like, “List household practices for conserving water
5. Reflective: continuous evaluation and their potential impact.”
and improvement of prompts Make a strategy request(…). Evaluate also the relevance and
applicability of each strategy. Consider the target audience's
needs and use this information to tailor future prompts to
generate content that better addresses specific challenges or
contexts.
https://doi.org/10.1016/j.acalib.2023.102720
3. Use of LLMs in the academic
workflow
The Research Process aided by AI (II)
with Responsible Conduct of Research at its core and some examples
- Grant writing
- Facilitate discovery and FAIR-ness - Research planning
- Explore open datasets - Study design
- Identify research gaps - Browsing and summarising the literature

- Write and editing

- Improve research outputs
(tables, figures)
- Peer review suggestions
- Interfaces for data collection

RCR
- Summarising for public
- Media reachout - Code and automation help

Ethics & Law

- Metadata preparation
- Summarising - Brainstorming on approaches for data
- Organise and store (meta)data preprocessing
- Code and automation
- Data minimization and annotation

- Annotations
- Methodological considerations
- Data analysis (Code and Original image from https://rdmkit.elixir-europe.org/ “The data lifecycle”
automation)
What approach
should I use?
Prompt a prompt
Coding (I) Step-by-step
Response
(this answer type
can be prompted)
Coding (II) Second derivative?

The slope does

not look like as
requested!
Coding (III) Ask for review!
Long forgotten
formulas
Name Grade1 Grade2 Grade3 Grade4 Average
--------- ------- ------- ------- ------- ---------
Alice 4.2 3.5 1.0 4.7 3.8
Bob 2.9 Text 4.2 5.1 #N/A
Charlie 3.7 4.0 3.2 6.3 #N/A
David 1.5 2.8 3.9 4.1 3.075
Emma 4.8 3.6 2.7 4.9 4.25
Frank 3.3 2.4 NaN 3.9 #N/A
Grace 4.7 2.1 3.9 5.2 3.975
Hannah 2.6 3.9 1.8 4.5 3.2
Ian 5.3 3.7 3.4 4.0 #N/A
Jack 3.1 4.5 3.6 2.7 3.475
Text review (I)
Text review (II)
4. Limitations, ethical and legal
considerations
Limitations & risks of (generative) AI
• Misinformation
• You will experience hallucinations confabulations
• LLMs cannot reason, but they can produce output that can
be meaningful to you (Enrico’s attempt with the classic example of the goat, wolf, and cabbage riddle)
• Reinforcement of biases (gender, racial, cultural, …)
• They are designed to output what has most frequently been
seen, not what should be the correct answer to your question
• Lack of reliable sources
• LLMs are not search engines, they are word generators
• Safeguards are in place, but they can be bypassed.
Responsible conduct of research
• Allea principles
• Reliability in ensuring the quality of research, reflected in the design, methodology,
analysis, and use of resources.
• Honesty in developing, undertaking, reviewing, reporting, and communicating research
in a transparent, fair, full, and unbiased way.
• Respect for colleagues, research participants, research subjects, society, ecosystems,
cultural heritage, and the environment.
• Accountability for the research from idea to publication, for its management and
organisation, for training, supervision, and mentoring, and for its wider societal impacts.

• TENK definition of research misconduct

• Fabrication (false data)
• Falsification (false results)
• Plagiarism (reusing without citing the sources)
Nothing new here, actually!
• AI cannot be given authorship
• Use of AI is should be transparent by describing how AI is used
• Personal data protection (GDPR) is to be followed
• Protect your own unpublished work
• The output of the AI system can be sensitive
• When creating artistic output, contributor roles must be explained
transparently and specifically.
• Reproducibility can be difficult to achieve
• Accountability: you are responsible for what you present
Aalto guidelines: https://www.aalto.fi/en/services/responsible-
use-of-artificial-intelligence-in-the-research-process
Be aware of the AI Hype
• You have most likely experienced the AI Hype
• grandiose narratives
• sparks of AGI
• AI scaremongering
• We are not going to end up like in the movie The Matrix
• The low-barrier for creating all sorts of AI-LLMs applications also
means lots of uncertain on where your personal and intellectual data
will go
Nothing new here, it is just another technology for researchers
and we need to evaluate if it can be useful for our work.
AI governance, AI ethics
• European Commission to soon approve the AI act: a framework for
risk evaluation of AI systems
• AI governance is what governments and companies are right now
struggling with (see CAIDP’s AI governance recommendations for the UN)
• As researchers we have more freedom of experimenting with these
new technologies, within the boundaries of responsible conduct
of research
• Aalto ethics committee recommends having an external ethics
advisor for AI. FCAI can help you find one.
5. Advanced cases and other
tools
Some more advanced uses of AI
1. Not just chat: all other sorts of NLP Natural Language Processing
2. Local open source LLMs for increased privacy, safety, and
reproducibility
3. Develop your AI application (paid APIs, local fine-tuning with OSS)
4. Existing external applications for academics
5. ChatGPT-4 plugins and other OpenAI advanced features
6. Beyond text: Midjourney, Dall.e, …
1. Language models beyond chats
• Not just chat-bots! Language model are useful for example for
automatic transcription of interviews.
• OpenAI Whisper is available on our Triton HPC cluster and
SECDATA environment, useful when interviews should not leave
Aalto storage. See the speech2text app developed by Aalto RSEs.
• Other uses are Named Entity Recognition for automatic
anonymisation of text, or to identify when a text is sensitive
• Neural Machine Translation with open-source models such as No
Language Left Behind (by Meta)
2. Local open source LLMs
• Open-source models
• You can download them and run them on your laptop
• Also available on Aalto Triton HPC cluster
• Personal favourite: model Llama2 (and CodeLlama) and
llamacpp for inference.
• Why local?
• Cloud scenario: you need Internet and data leaves you
• Your Computer -> ChatGPT website -> remote inference software -> remote
model
• Local scenario: no internet needed, data stays with you
• Your computer (or Aalto Triton HPC) -> local terminal/website -> local inference
software -> local model
3. Build your AI app
• Build a chatbot with Azure OpenAI API
• Or build it locally using llamacpp + open-source models +
development tools such as LangChain, LangSmith, PineCone
• Fine tuning
• Prepare a new set of instructions (e.g. questions and
answers)
• Tweak the model weights so that the model is fine-tuned with
the new content
• It can be computationally expensive (many GPUs), but there
are efficient ways with few compromises (qLORA)
4. External AI applications
• Here a list of Apps that can be useful for academics
• ChatPDF
• Dante.ai
• Cursor
• GitHub copilot
• ...what is your favourite?
5. ChatGPT-4 new features
• For 20 USD/month you can access GPT4 with Beta features
• Custom instructions
• Advanced data analysis
• Better model
• Plugins
• Beware with plugins since it is not always clear where your data will go
6. Beyond text
1. DALL·E 3 (openai.com)
2. Midjourney
3. Replicate
6. Conclusions and references
Take home message
• AI is a junior assistant you can hire: you don’t maybe want to
assign the most critical tasks to them, you want to assign tedious
tasks where you are an expert so that it’s easy to verify the output or
assign exploratory tasks for you later to evaluate
• Autoregressive LLMs as they are today are not “intelligent”. They
are great at generating an average of all the things they were
exposed to…. And carry all the biases and errors from the past
• Responsible Conduct of Research is at the core of any academic
activity that you do: be honest and transparent
Where to go from here?
• There are great resources for self learning, beware of scammers.
A reliable list available at this link
• Aalto researchers: are you unsure of the possibilities for your
research? Talk with us: [email protected] is a network of
networks that works closely with IT services, Ethics, Legal, Software
Development Services.
• Teachers: there are similar considerations for teachers, but
protecting student data is of absolute importance. It would
require a dedicated training session; most likely you are already
discussing this with your colleagues.
References
• Unesco guidelines for generative AI
• Excellent guide from Aarhus University
• Jeremy Howards's Hacker guide to Language
Models
• Elements of AI open course
• Ethics of AI open course

A list of questions and answer from the live training session is available at https://www.aalto.fi/en/open-science-and-research/ai-and-
research-work-useful-learning-materials .
 Thank you!

aalto.fi/en/open-science-and-research