Authenticity and Assessment �in the Age of AI

Dominik.Lukes@ctl.ox.ac.uk

Reading & Writing Innovation Lab

bit.ly/ox-rewrilab Consultations for staff and students

Beyond ChatGPT: State of AI, October 2023

Download slides

bit.ly/3yOlMAP

Language pedagogy is my passion

The three As

Authenticity

Assessment

AI

Three pairs

Authenticity

Assessment

AI

Authenticity

Assessment

AI

Assessment and authenticity

Assessment has to solve more than authenticity.

Assessment has to be 3 things at once

But let’s remember the engineering triangle

Pick any two

Good

Fast

Cheap

Engineering triangle in practice

Do we have an assessment triangle?

Pick any two

Authentic

Scalable

Reliable

Assessment triangle in practice

Note: Not my original idea. Overheard at a conference.

Cognitive break and questions

In the triangle of assessment, we try,

To be scalable, reliable, and fly,

Yet authenticity’s claim,

Is a part of the game,

Balancing three is the aim, oh my!

How I made the previous slide

AI and authenticity

What is AI?�Brief terminological detour

Often heard

ChatGPT is just a fancy autocomplete. It just predicts the next word.

Literally true, but not a useful way to think about AI.

Generative AI is a universal (semantic) translator

Translating languages, styles, modalities

Some things don’t feel like translation but AI treats them that way

This approach can even be used to replace old specialised systems

But why semantic? Doesn’t it just predict the next word?

The relations look something like this

When AI is generating text it is doing something like this

This may look simple in 3D space

What is inside an LLM is not statistics but geometry.

BUT it’s all geometry in 10,000 dimensions

Each item is one giant vector

[-0.08041892945766449, -0.023566607385873795, 0.04585130512714386, 0.037420596927404404, -0.09120217710733414, 0.022545181214809418, -0.0019880179315805435, 0.0587424635887146, 0.07126272469758987, -0.02159898355603218, -0.07145281136035919, 0.09984641522169113, -0.05501342564821243, 0.02485564909875393, 0.01755008101463318, 0.014556304551661015, -0.15110555291175842, -0.000567720562685281, 0.10030079632997513, -0.045505933463573456, -0.06274029612541199, -0.0683555155992508, 0.0008911662152968347, 0.01842709816992283, 0.06299598515033722, 0.02255615033209324, -0.09917508065700531, -0.07070962339639664, 0.08635025471448898, 0.06686452776193619, -0.0407336950302124, -0.04072333127260208, -0.01974628120660782, 0.07472220063209534, -0.024722471833229065, -0.13420116901397705, -0.01812688820064068, -0.07096941769123077, -0.05353084206581116, -0.10960721969604492, -0.017906684428453445, -0.04733094945549965, -0.02091103047132492, 0.1269848346710205, -0.05413510650396347, -0.046787846833467484, 0.0024005023296922445, -0.07217800617218018, -0.029329143464565277, 0.007498500403016806, -0.034666456282138824, -0.03568940982222557, 0.03427724167704582, 0.02315753698348999, -0.008645392023026943, 0.05333952233195305, 0.07456360012292862, 0.147796630859375, -0.006483903620392084, -0.08905889838933945, 0.03265034034848213, -0.0732979029417038, 0.04066538065671921, 0.023211032152175903, -0.012049349024891853, -0.02828565053641796, 0.019329581409692764, 0.09989447146654129, 0.1430598795413971, -0.061100199818611145, -0.030345138162374496, -0.02984507940709591, -0.028366880491375923, 0.052052564918994904, 0.036766957491636276, 0.003982939291745424, -0.077084481716156, 0.05044832453131676, -0.11687757074832916, 0.06646141409873962, 0.016255078837275505, -0.06982151418924332, -0.000822143629193306, -0.0026820336934179068, -0.004263593349605799, 0.09659365564584732, 0.06130471080541611, -0.06840908527374268, 0.06686245650053024, -0.04831290245056152, 0.08598440140485764, 0.08331689983606339, 0.08026000112295151, 0.05451888591051102, -0.03798443824052811, 0.04084145650267601, -0.12311697751283646, 0.023645302280783653, 0.005237551871687174, 0.03906212002038956, 0.037468183785676956, -0.05121520534157753, -0.10456130653619766, 0.009842721745371819, 0.04819759353995323, -0.13286681473255157, 0.02991127222776413, -0.06024811416864395, 0.04108288511633873, -0.008447377011179924, -0.07916080206632614, 0.06436653435230255, 0.017831943929195404, -0.054629500955343246, 0.027066148817539215, -0.030593710020184517, -0.10156133025884628, -0.0013401528121903539, 0.0011191506637260318, 0.009616676717996597, -0.02962290495634079, 0.0042936066165566444, 0.013841508887708187, -0.047656722366809845, -0.003912750165909529, 0.06500802934169769, 0.001283025718294084, -0.0816996768116951, 0.06566621363162994, -0.010957532562315464, -0.028156422078609467, 0.08978854864835739, -0.0003194105520378798, -0.02697799727320671, -0.006005867850035429, 0.07932088524103165, 0.021490609273314476, 0.013727870769798756, -0.019940776750445366, 0.031798265874385834, -0.0457642637193203, 0.03235720098018646, -0.022082772105932236, -0.04902353510260582, -0.11819718778133392, -0.04506421089172363, -0.046244461089372635, 0.029877550899982452, -0.07711911201477051, 0.05314543470740318, -0.09000932425260544, -0.023750705644488335, -0.05107633396983147, 0.001467616413719952, -0.02442317083477974, 0.01248782780021429, 0.06548482179641724, 0.043813593685626984, 0.06102786585688591, 0.021692050620913506, -0.052160654217004776, -0.009674523957073689, -0.072069451212883, -0.08633119612932205, -0.05121589079499245, -0.08108754456043243, 0.03608304262161255, 0.06553766876459122, -0.0727415531873703, -0.09346839785575867, -0.07251054048538208, 0.04504929482936859, -0.01773262582719326, -0.0005254627903923392, -0.0035706141497939825, 0.09068302065134048, 0.0152428038418293, 0.009525319561362267, 0.02502918615937233, 0.02807294949889183, -0.08951258659362793, 0.018022941425442696, 0.04113161191344261, -0.09941867738962173, 0.03642140328884125, 0.07755865901708603, 0.014834643341600895, -0.05757498741149902, -0.0052739898674190044, -0.03217893838882446, 0.029460914433002472, -0.03587955981492996, 0.016881171613931656, -0.015574142336845398, -0.10131996870040894, -0.01736866682767868, 0.014807181432843208, -0.03830776736140251, -0.0307577196508646, -0.04063287377357483, 0.0017508701421320438, 0.06622152030467987, 0.06959225982427597, 0.03921446576714516, -0.029292205348610878, -0.07731080055236816, -0.0757351890206337, 0.008267058990895748, 0.10628201067447662, -0.006961626932024956, -0.060704007744789124, -0.024280674755573273, -0.011232278309762478, 0.02305467799305916, -0.040246833115816116, 0.03551888465881348, -0.12048669904470444, -0.0057440041564404964, -0.008801680989563465, -0.038733456283807755, -0.0384967215359211, -0.0059003811329603195, 0.07543318718671799, 0.0029998512472957373, 0.11148137599229813, 0.0560586079955101, -0.01694066822528839, -0.020253779366612434, -0.11995487660169601, 0.10403268039226532, -0.022030610591173172, 0.019188301637768745, -0.03581297770142555, -0.04047590494155884, -0.03492145985364914, 0.027967417612671852, -0.07497915625572205, 0.032431814819574356, -0.025854842737317085, -0.10595495998859406, -0.09982465207576752, -0.05515384301543236, 0.02156943641602993, 0.05118619278073311, -0.03904290497303009, -0.022826874628663063, -0.053247928619384766, -0.10935184359550476, 0.0006719367229379714, -0.016026955097913742, 0.13483813405036926, 0.1173691526055336, -0.01902260072529316, -0.09690848737955093, -0.07585378736257553, 0.007626112550497055, 0.019889818504452705, -0.008633404038846493, 0.010355712845921516, 0.035737670958042145, 0.011519350111484528, -0.005264237057417631, -0.06305427849292755, -0.026263760402798653, 0.008310412988066673, -0.0068666874431073666, -0.13443514704704285, -0.025350390002131462, -0.0079041114076972, 0.014966381713747978, 0.01571144163608551, 0.06266333907842636, 0.05788900703191757, -0.022854981943964958, 0.09513315558433533, 0.1284472942352295, -0.061813995242118835, -0.049407169222831726, -0.10701776295900345, 0.06945358961820602, -0.07409369200468063, -0.028664348646998405, -0.0144350565969944, 0.029182329773902893, 0.007034373469650745, -0.026693496853113174, 0.0590004064142704, -0.002902168082073331, 0.12047384679317474, 0.023063501343131065, -0.05780957639217377, 0.058589596301317215, 0.02074800431728363, -0.030389118939638138, -0.002812192542478442, 0.06409497559070587, -0.0015993582783266902, 0.007702010218054056, 0.013223372399806976, 0.012501182034611702]

This gives AI enormous power but it has limitations.

Big 3 AI limitations

Biggest caveat

It is not always possible to tell ahead of time what AI will be good at.

Dell’Acqua et al. 2023: Navigating the Jagged Technological Frontier

“We suggest that the capabilities of AI create a “jagged technological frontier” where some tasks are easily done by AI, while others, though seemingly similar in difficulty level, are outside the current capability of AI.”

Can we tell how good AI would be at something?

Illustrating the Jagged Frontier - @techczech

“We suggest that the capabilities of AI create a “jagged technological frontier” where some tasks are easily done by AI, while others, though seemingly similar in difficulty level, are outside the current capability of AI.”

Dell’Acqua et al. 2023: Navigating the Jagged Technological Frontier

Metalanguage fail example (Claude Opus 17 May 2024)

How people experience the jagged frontier @techczech

“We suggest that the capabilities of AI create a “jagged technological frontier” where some tasks are easily done by AI, while others, though seemingly similar in difficulty level, are outside the current capability of AI.”

Dell’Acqua et al. 2023: Navigating the Jagged Technological Frontier

😍

ChatGPT is amazing, there’s nothing it can’t do!

😡

ChatGPT is useless, it can’t get even the basics right!

Most tasks sit at an intersection.

AI is good as expected

AI is �unexpectedly bad

AI is much better than expected

AI cannot do this� as expected

A conversation I’ve had

Academic: Why are the references ChatGPT gives me wrong?

Me: ChatGPT hallucinates links and references. Do not use it to find them.

A conversation I’ve had

Me: ChatGPT hallucinates links and references. Don’t use it to find them.�Student: Every time I clicked on a link I asked for it worked.

Walters & Wilder 2023 Fabrication and errors in the bibliographic citations generated by ChatGPT

AI has its own assessment triangle problem

All AI tools are powered by models.

What is a model

One model many tools

Different kinds of models

More than a few

Which ones are best?

ChatGPT vs ChatGPT Plus

Difference between GPT3.5 and GPT-4

Example of improvements

GPT-3.5 is also less powerful in language

Spectrum of capabilities

Three types of model assessment

Benchmarks

MMLU: Most popular benchmark

Top models on MMLU in May 2024

But other benchmarks give different scores

Head-to-head

A good place to learn about them is LMSys Arena

Recent leaderboard – April 2024

Vibes

Let’s keep this in mind

How do you know what is on which side of the jagged frontier?

What can you do to get a “feel” for the AI jagged frontier?

Cognitive break and questions

Why did the robot chAIcken cross the road?

​

To get to the jagged technological frontier—because speaking any language is a breeze, but counting words in a paragraph? Now that's a real challenge!

AI and Assessment

What is the new ‘authenticity’ of the tasks we assess?

There’s an authenticity lag in assessment

Often heard

Spelling does not matter in the age of the spell check!

Spelling does not matter

Spelling becomes reliable indicator of general knowledge and skill

What is authentic to core academic practice?

The future of writing is conversation?

Dictation vs listening

👎 Dictation

Listening 👍

You are not dictating. You are sharing your thoughts.

Conversation

👎 Dictation

Conversation 👍

The future of writing is conversation

What this means: New form of writing

Output

How does reading change with AI cognitive scaffolding?

Step 1: �Ask for bullets

Step 2: �Ask for propositions

Step 3: �Ask for tables

Step 4: �Ask for questions

Step 5: �Ask for examples

Step 6: �Ask for poetry

Cognitive break and questions

AI shapes learning,

Authentic tasks redefine,

Future in our hands.

Final dilemma

Google announced watermarkign

Should AI companies watermark all text generated by AI?

Should spelling checkers send information about what errors you made to the company who will give a list to your employer?

Questions

Thank you

Dominik.Lukes@ctl.ox.ac.uk @techczech

This presentation is licensed under Creative Commons By Attribution license except where otherwise noted.��Icons and stock photos licensed under Microsoft Premium Content and cannot be reused outside this document.