Intelligent Systems Podcast: Andrew Olney on AI, Language, Tutoring, and the Philip K. Dick Android

About Dr. Andrew Olney

Dr. Andrew Olney is a pro­fes­sor in both the Insti­tute for Intel­li­gent Sys­tems and the Depart­ment of Psy­chol­o­gy at the Uni­ver­si­ty of Mem­phis. His research brings togeth­er arti­fi­cial intel­li­gence, nat­ur­al lan­guage pro­cess­ing, edu­ca­tion, psy­chol­o­gy, inter­faces, and con­ver­sa­tion­al learn­ing tech­nolo­gies.

Olney earned a bachelor’s degree in lin­guis­tics and cog­ni­tive sci­ence from Uni­ver­si­ty Col­lege Lon­don, a master’s degree in evo­lu­tion­ary and adap­tive sys­tems from the Uni­ver­si­ty of Sus­sex, and a PhD in com­put­er sci­ence from the Uni­ver­si­ty of Mem­phis. His work has focused espe­cial­ly on using lan­guage tech­nolo­gies to assess and improve learn­ing from text and learn­ing through con­ver­sa­tion.

From Early Computers to AI Research

Olney describes a life­long inter­est in com­put­ers and arti­fi­cial intel­li­gence. His father worked as a sys­tems engi­neer for IBM, so com­put­ers were a reg­u­lar part of his home life. He began using an Apple II Plus as a young child and start­ed pro­gram­ming in BASIC at around age sev­en.

As he pro­gressed through high school and col­lege, his inter­ests expand­ed from pro­gram­ming into lan­guage, psy­chol­o­gy, and the rela­tion­ship between lan­guage and thought. This led him to study lin­guis­tics and cog­ni­tive sci­ence, fol­lowed by work on neur­al net­works and nat­ur­al lan­guage pro­cess­ing.

Dur­ing his master’s work, Olney explored evolv­ing neur­al net­works. How­ev­er, the com­put­ing pow­er and data avail­able at the time lim­it­ed what those sys­tems could accom­plish. He lat­er returned to the Unit­ed States, con­nect­ed with the Insti­tute for Intel­li­gent Sys­tems through a con­sult­ing oppor­tu­ni­ty, and even­tu­al­ly pur­sued a PhD in com­put­er sci­ence while work­ing on intel­li­gent tutor­ing research with Dr. Art Graess­er.

What Makes a System Intelligent?

Olney explains that defin­ing intel­li­gence is dif­fi­cult, much like defin­ing life. There may not be a sin­gle point at which some­thing sud­den­ly becomes intel­li­gent. Instead, intel­li­gence can be under­stood as a con­tin­u­um.

Ques­tions about intel­li­gence become more com­pli­cat­ed as the exam­ples change: Is a dog intel­li­gent? A worm? A bac­teri­um? A ther­mo­stat? A light switch? The answer often depends on the capa­bil­i­ties being con­sid­ered and the con­text in which a sys­tem is oper­at­ing.

Olney is par­tic­u­lar­ly inter­est­ed in AI sys­tems that can per­form tasks at lev­els com­pa­ra­ble to human per­for­mance. He notes, how­ev­er, that a sys­tem can appear intel­li­gent with­in a nar­row con­text with­out pos­sess­ing broad intel­li­gence.

One his­tor­i­cal exam­ple is ELIZA, an ear­ly chat­bot that imi­tat­ed a Roger­ian psy­chother­a­pist by reflect­ing a user’s state­ments back to them. Although ELIZA used a rel­a­tive­ly sim­ple approach, its fram­ing made it feel con­vinc­ing to some users because reflec­tive respons­es fit the expec­ta­tions of that ther­a­peu­tic set­ting.

Intelligent Tutoring Systems

Olney describes intel­li­gent tutor­ing sys­tems as AI sys­tems that take on some of the func­tions of a human tutor. These sys­tems have instruc­tion­al objec­tives, hold con­ver­sa­tions with stu­dents, esti­mate what a stu­dent knows, and adapt their teach­ing strat­e­gy based on that assess­ment.

Ear­li­er tutor­ing sys­tems used tech­niques such as latent seman­tic analy­sis, or LSA, to rep­re­sent lan­guage math­e­mat­i­cal­ly. A student’s response could be con­vert­ed into a vec­tor rep­re­sen­ta­tion and com­pared with expect­ed answers or rel­e­vant con­cepts.

The sys­tem used this infor­ma­tion to cre­ate a work­ing mod­el of the learner’s knowl­edge. It could ask diag­nos­tic ques­tions, esti­mate the student’s con­fi­dence or under­stand­ing of dif­fer­ent con­cepts, and focus instruc­tion on mate­r­i­al the stu­dent was less like­ly to under­stand.

In oth­er words, the sys­tem did not mere­ly deliv­er the same les­son to every stu­dent. It attempt­ed to deter­mine what each learn­er already knew and then adjust the con­ver­sa­tion and instruc­tion accord­ing­ly.

Can AI Learn From Users?

Olney dis­tin­guish­es between mul­ti­ple forms of learn­ing in old­er intel­li­gent tutor­ing sys­tems. One kind of learn­ing hap­pened before a stu­dent ever used the sys­tem: lan­guage mod­els were built from source mate­r­i­al so the tutor could inter­pret stu­dent respons­es.

Anoth­er kind hap­pened dur­ing inter­ac­tion. The sys­tem learned about the stu­dent by ana­lyz­ing respons­es, test­ing under­stand­ing through ques­tions, and updat­ing its esti­mate of the student’s knowl­edge.

Olney also exper­i­ment­ed with sys­tems that could learn what they were sup­posed to teach while inter­act­ing with users. This “self-boot­strap­ping” approach did not work espe­cial­ly well in prac­tice because most users did not want to spend the addi­tion­al effort need­ed to train or cor­rect the sys­tem. They expect­ed the tool to work well imme­di­ate­ly.

He con­nects this find­ing to mod­ern large lan­guage mod­els: users may notice errors in AI out­put, but only a rel­a­tive­ly small num­ber are will­ing to invest the time need­ed to pro­vide detailed cor­rec­tions or help improve the sys­tem.

Language, Mirroring, and Rapport

The con­ver­sa­tion also explores whether tutor­ing sys­tems should mir­ror the syn­tax, lan­guage pat­terns, pos­ture, or non­ver­bal behav­ior of the peo­ple using them.

Olney’s research includ­ed ani­mat­ed tutor agents that could use syn­the­sized speech, ges­tures, facial expres­sions, and oth­er non­ver­bal cues. Some ver­sions exper­i­ment­ed with non­ver­bal mir­ror­ing, such as hav­ing an agent lean for­ward when a stu­dent leaned for­ward or smile when a stu­dent smiled.

He also dis­cuss­es research on syn­tac­tic prim­ing and lin­guis­tic align­ment. Peo­ple can rapid­ly begin using lan­guage struc­tures they have recent­ly heard, and these effects can some­times per­sist beyond the imme­di­ate inter­ac­tion.

Mir­ror­ing may help cre­ate rap­port and encour­age learn­ers to return to a sys­tem. How­ev­er, Olney notes that it is less clear whether mir­ror­ing direct­ly improves what a learn­er under­stands or retains in the moment.

Where Does the System End?

Olney argues that the “sys­tem” part of intel­li­gent sys­tems is also dif­fi­cult to define. An indi­vid­ual ani­mal may be treat­ed as a sys­tem, but groups of ani­mals can also behave as sys­tems.

For exam­ple, a flock of birds or school of fish can pro­duce coor­di­nat­ed, emer­gent behav­ior even when each indi­vid­ual fol­lows rel­a­tive­ly sim­ple local rules. Sim­i­lar­ly, a per­son using a com­put­er or anoth­er tool can be under­stood as part of a larg­er human-tech­nol­o­gy sys­tem.

For Olney, the most impor­tant sci­en­tif­ic ques­tion is whether defin­ing a col­lec­tion of peo­ple, tools, or organ­isms as a sys­tem pro­vides use­ful explana­to­ry pow­er. If the sys­tem per­spec­tive helps researchers make pre­dic­tions or under­stand behav­ior that would oth­er­wise be dif­fi­cult to explain, then it is valu­able.

The Philip K. Dick Android

One of the most mem­o­rable parts of the episode is Olney’s account of the Philip K. Dick Android, an inter­ac­tive robot­ic head cre­at­ed in col­lab­o­ra­tion with roboti­cist and artist David Han­son.

Han­son was known for high­ly real­is­tic robot­ic facial sculp­tures and was inter­est­ed in com­bin­ing life­like ani­ma­tron­ics with arti­fi­cial intel­li­gence. Olney’s back­ground in Auto­Tu­tor, nat­ur­al lan­guage pro­cess­ing, speech syn­the­sis, and con­ver­sa­tion­al sys­tems made the col­lab­o­ra­tion a nat­ur­al fit.

The project used mate­r­i­al from Philip K. Dick’s work and life, includ­ing scanned nov­els and pub­lished con­ver­sa­tions with the sci­ence-fic­tion author. Olney cre­at­ed mul­ti­ple con­ver­sa­tion­al com­po­nents for the Android, includ­ing:

• Respons­es drawn from pub­lished con­ver­sa­tions with Philip K. Dick
• A chat­bot for com­mon ques­tions peo­ple ask robots
• Lan­guage mod­els built from Philip K. Dick’s books
• A sys­tem for select­ing the most appro­pri­ate response in real time

The Android was demon­strat­ed at Wired Nextfest, received nation­al media atten­tion, and lat­er won an open-inter­ac­tion robot­ics com­pe­ti­tion at the Asso­ci­a­tion for the Advance­ment of Arti­fi­cial Intel­li­gence con­fer­ence.

The orig­i­nal robot­ic head was lat­er lost dur­ing air trav­el while being trans­port­ed to a pre­sen­ta­tion at Google in Moun­tain View. Olney jokes that it may still be sit­ting in an air­port stor­age facil­i­ty, wait­ing to be redis­cov­ered cen­turies in the future.

What People Miss About AI

Olney believes that many pub­lic dis­cus­sions of AI focus too nar­row­ly on the lat­est vis­i­ble tech­nol­o­gy, espe­cial­ly large lan­guage mod­els. He argues that AI has a much longer his­to­ry and includes many ideas, meth­ods, and appli­ca­tions beyond sys­tems such as Chat­G­PT.

Large lan­guage mod­els are impres­sive and use­ful, but they are not the entire­ty of arti­fi­cial intel­li­gence. They have strengths and lim­i­ta­tions, and they are only one part of a much broad­er tech­ni­cal and intel­lec­tu­al field.

Olney com­pares this moment to the ear­ly devel­op­ment of mag­net­ic stor­age. The vis­i­ble appli­ca­tion, such as an answer­ing machine, may cap­ture pub­lic atten­tion, while the under­ly­ing tech­nol­o­gy is what even­tu­al­ly cre­ates much broad­er and more last­ing change.

In the same way, he sug­gests that peo­ple should look beyond the imme­di­ate nov­el­ty of con­ver­sa­tion­al AI and con­sid­er the under­ly­ing advances that may shape future sys­tems, tools, and forms of human-com­put­er inter­ac­tion.

Looking Ahead

As he approach­es retire­ment, Olney says he remains focused on com­plet­ing research projects and writ­ing up old­er work that has not yet been pub­lished. Rather than slow­ing down, he is con­cen­trat­ing on fin­ish­ing as much mean­ing­ful work as pos­si­ble before leav­ing the Uni­ver­si­ty of Mem­phis.

The episode high­lights a career shaped by curios­i­ty about com­put­ers, lan­guage, learn­ing, and human intel­li­gence. From ear­ly tutor­ing sys­tems to embod­ied con­ver­sa­tion­al agents and robot­ic inter­faces, Olney’s work shows how AI can be used not only to auto­mate tasks but also to study how peo­ple think, com­mu­ni­cate, and learn.