MIS Working with AI Review

When did AI first start being developed? What year?

in the 1950s, in 1956 AI becomes a field of study

What does traditional AI do?

classifies or predicts outcomes from historical data

What are some examples of traditional AI?

• spam filters

• fraud detection

• sales and demand forecasts

• customer segmentation

When should you use traditional AI?

• the data is private or sensitive

• the takes needs niche expert knowledge

• a working model already exists

Traditional AI cannot…

cannot create anything new

What are the 3 steps for traditional AI?

1. historical data

2. finds patterns

3. prediction or category

REACTIVE: ANALYZES and PREDICTS

What does generative AI do?

creates new content: text, images, video, and code

What are some examples of generative AI?

• chatbots and conversational search

• summaries and translation

• marketing content and images

• writing and explaining code

When should you use generative AI?

• the take uses everyday language or common images

• you need it running fast, with less technical expertise

What should you watch out for with generative AI?

inaccuracies and bias

What are the 3 steps to generative AI?

1. prompt

2. learned patterns

3. new content

PROACTIVE: CREATES SOMETHING NEW

What are traditional AI and generative AI best used together?

generative AI can give a machine learning model more context, and can help build one

What are the 7 behaviors you can see from AI?

1. Variability

2. Hallucination

3. Bias

4. Knowledge cutoff

5. Context limits

6. Sycophancy

7. Opacity

What is variability?

the same question gets different answers

Why does next-word prediction cause variability?

it picks likely words with some randomness, so each answer can come out a little different

What is hallucination?

it makes something up and says it with confidence

Why does next-word prediction cause hallucination?

a fake source can sound just as likely as a real one, confidence is style, not proof

What is bias?

unfair patterns, like stories where the executive is always a man

Why does next-word prediction cause bias?

it learned its patterns from training data, including the unfair ones

What is knowledge cutoff?

It does not know recent events

Why does next-word prediction cause knowledge cutoff?

its training stopped at a data, it can only predict what it saw

What is context limits?

long chats get forgotten

Why does next-word prediction cause context limits?

it can only hold so much of one conversation at a time

What is sycophancy?

it agrees with you just because you said it, even when you are wrong

Why does next-word prediction cause sycophancy?

agreeing with the person often looks like the most likely next answer

What is opacity?

no one can fully explain why it gave a specific answer

Why does next-word prediction cause opacity?

the answer comes from billions of patterns, not from steps anyone can read

What should you avoid sharing with AI?

• other people’s personal information

• grade or records

• company confidential data

• passwords

What is prompt injection?

hidden text in a file or website that tries to give the AI orders

AI can have trouble telling your instructions apart from instructions hidden inside a document

Why does AI slop happen?

slop is not the AI being lazy, it is the AAI doing exactly what it does, with too little to go on

What are the signs of AI slop?

1. No specifics - it could be about any company

2. Filler phrases - openers like “in today’s fast-paced digital world”

3. Fake enthusiasm - words like “exciting” or “game-changing” with nothing to back them up

4. Repeats the question - it restates what you asked instead of answering it

5. Sounds like everyone - it reads like very other AI answer you have seen

How do you fix AI slop?

give the AI a business profile, set constraints, add guardrails, and add your own insight

What does data mean for AI?

give it the source

What does constraints mean for AI?

who is it for, how long, what format (ex: for a manager, exactly 5 bullets, plain language)

What does guardrails mean for AI?

what not to do (ex: “if it’s not i the profile write not in source”)