The future of brain-machine synchronisation

The remark­able evo­lu­tion of Arti­fi­cial Intel­li­gence (AI) sys­tems rep­re­sents a par­a­digm shift in the rela­tion­ship between humans and machines. This trans­for­ma­tion is evi­dent in the seam­less inter­ac­tions facil­i­tat­ed by these advanced sys­tems, where adapt­abil­i­ty emerges as a defin­ing char­ac­ter­is­tic, res­onat­ing with the fun­da­men­tal human capac­i­ty to learn from expe­ri­ence and pre­dict behaviour.

One facet of AI that aligns close­ly with human cog­ni­tive process­es is Rein­force­ment Learn­ing (RL). RL mim­ics the human learn­ing par­a­digm by allow­ing AI sys­tems to learn through inter­ac­tion with an envi­ron­ment, receiv­ing feed­back in the form of rewards or penal­ties. By con­trast, Large Lan­guage Mod­els (LLMs) play a cru­cial role in pat­tern recog­ni­tion, cap­tur­ing the intri­cate nuances of human lan­guage and behav­iour. These mod­els, such as Chat­G­PT and BERT, excel in under­stand­ing con­tex­tu­al infor­ma­tion, grasp­ing the sub­tleties of lan­guage, and pre­dict­ing user intent. Lever­ag­ing vast datasets, LLMs acquire a com­pre­hen­sive under­stand­ing of lin­guis­tic pat­terns, enabling them to gen­er­ate human-like respons­es and adapt to some of the user behav­iour, some­times with remark­able accuracy.

The syn­er­gy between RL and LLMs cre­ates a pow­er­ful pre­dic­tor of human behav­iour. RL con­tributes the abil­i­ty to learn from inter­ac­tions and adapt, while LLMs enhance the pre­dic­tion capa­bil­i­ties through pat­tern recog­ni­tion. AI sys­tems based on RL can thus dis­play a form of behav­iour­al syn­chrony. At its core, RL enables AI sys­tems to learn opti­mal sequences of actions in inter­ac­tive envi­ron­ments to achieve a pol­i­cy. Anal­o­gous to a child touch­ing a hot sur­face and learn­ing to avoid it, these AI sys­tems adapt based on the pos­i­tive or neg­a­tive feed­back they receive.

AI agents using deep rein­force­ment learn­ing, such as Google Deep­Mind’s Alp­haZe­ro, learn and improve by play­ing mil­lions of games against them­selves, there­by refin­ing their strate­gies over time. This self-improve­ment process in AI involves an agent iter­a­tive­ly learn­ing from its own actions and out­comes. Sim­i­lar­ly, in human inter­ac­tions, brain syn­chrony occurs when indi­vid­u­als engage in coop­er­a­tive tasks, lead­ing to aligned pat­terns of brain activ­i­ty that facil­i­tate shared under­stand­ing and col­lab­o­ra­tion. Unlike AI, humans achieve this syn­chrony through inter­ac­tion with oth­ers rather than themselves.

What’s more, AI sys­tems can also learn from inter­ac­tions with humans. Just as human brain syn­chrony enhances coop­er­a­tion and under­stand­ing, AI sys­tems can improve and align their respons­es through exten­sive iter­a­tive learn­ing from human inter­ac­tions. While AI sys­tems do not lit­er­al­ly share knowl­edge as human brains do, they become repos­i­to­ries of data inher­it­ed from these inter­ac­tions, which cor­re­sponds to a form of knowl­edge. This process of learn­ing from vast datasets, includ­ing human inter­ac­tions, can be seen as a form of ‘col­lec­tive mem­o­ry’. This anal­o­gy high­lights the poten­tial for AI sys­tems to evolve while being influ­enced by humans, while also influ­enc­ing humans through their use, indi­cat­ing a form of ‘com­pu­ta­tion­al syn­chrony’ that could be seen as an ana­logue to human brain synchrony.

In addi­tion, AI sys­tems enabled with social cue recog­ni­tion are being designed to detect and respond to human emo­tions. These ‘Affec­tive Com­put­ing’ sys­tems, as coined by Ros­alind Picard in 1995¹, can inter­pret human facial expres­sions, voice mod­u­la­tions, and even text to gauge emo­tions and then respond accord­ing­ly. An AI assis­tant that can detect user frus­tra­tion in real-time and adjust its respons­es or assis­tance strat­e­gy is a rudi­men­ta­ry form of behav­iour­al syn­chro­ni­sa­tion based on imme­di­ate feedback.

For instance, affec­tive com­put­ing encom­pass­es tech­nolo­gies like emo­tion recog­ni­tion soft­ware that analy­ses facial expres­sions and voice tone to deter­mine a person’s emo­tion­al state. Real-time sen­ti­ment analy­sis in text and voice allows AI to adjust its inter­ac­tions to be more empa­thet­ic and effec­tive. This capa­bil­i­ty is increas­ing­ly used in cus­tomer ser­vice chat­bots and vir­tu­al assis­tants to improve user expe­ri­ence by mak­ing inter­ac­tions feel more nat­ur­al and responsive.

Just as humans adjust their behav­iour in response to social cues, adap­tive AI sys­tems mod­i­fy their actions based on user input, poten­tial­ly lead­ing to a form of ‘syn­chro­ni­sa­tion’ over time. Assess­ing the social com­pe­tence of such an AI sys­tem could be done by adapt­ing tools like the Social Respon­sive­ness Scale (SRS)—a well-val­i­dat­ed psy­chi­atric instru­ment that mea­sures how adept an indi­vid­ual is at mod­i­fy­ing their behav­iour to fit the behav­iour and dis­po­si­tion of a social part­ner, a proxy for ‘the­o­ry of mind,’ which refers to the abil­i­ty to attribute men­tal states—such as beliefs, intents, desires, emo­tions, and knowledge—to one­self and to others.

Brain-Com­put­er Inter­faces (BCIs) have ush­ered in a trans­for­ma­tive era in which thoughts can be trans­lat­ed into dig­i­tal com­mands and human com­mu­ni­ca­tion. Com­pa­nies like Neu­ralink are mak­ing strides devel­op­ing inter­faces that enable paral­ysed indi­vid­u­als to con­trol devices direct­ly with their thoughts. Con­nect­ing direct record­ings of brain activ­i­ty with AI sys­tems, researchers enabled an indi­vid­ual to speak at nor­mal con­ver­sa­tion­al speed after being mute for more than a decade fol­low­ing a stroke. AI sys­tems can also be used to decode not only what an indi­vid­ual is read­ing but what they are think­ing based on non-inva­sive mea­sures of brain activ­i­ty using func­tion­al MRI.

Based on these advances, it’s not far-fetched to imag­ine a future sce­nario in which a pro­fes­sion­al uses a non-inva­sive BCI (e.g., wear­able brain­wave mon­i­tors such as Cog­wear, Emo­tiv, or Muse) to com­mu­ni­cate with AI design soft­ware. The soft­ware, recog­nis­ing the designer’s neur­al pat­terns asso­ci­at­ed with cre­ativ­i­ty or dis­sat­is­fac­tion, could instan­ta­neous­ly adjust its design pro­pos­als, achiev­ing a lev­el of syn­chrony pre­vi­ous­ly thought to be the realm of sci­ence fic­tion. This tech­no­log­i­cal fron­tier holds the promise of a dis­tinc­tive form of syn­chrony, where the inter­play between the human brain and AI tran­scends mere com­mand inter­pre­ta­tion, open­ing up a future in which AI res­onates with human thoughts and emotions.

Cru­cial­ly, the res­o­nance envi­sioned here tran­scends the behav­iour­al domain to encom­pass com­mu­ni­ca­tion as well. As BCIs evolve, the poten­tial for out­ward expres­sions becomes piv­otal. Beyond mere com­mand exe­cu­tion, the inte­gra­tion of facial cues, tone of voice, and oth­er non-ver­bal cues into AI’s respons­es ampli­fies the chan­nels for res­o­nance. This expan­sion into mul­ti­modal com­mu­ni­ca­tion may enrich syn­chrony by cap­tur­ing ele­ments from the holis­tic nature of human expres­sion, cre­at­ing a more immer­sive and nat­ur­al interaction.

How­ev­er, the con­cept of res­o­nance also presents the chal­lenge of nav­i­gat­ing the uncan­ny val­ley, a phe­nom­e­non where humanoid enti­ties that close­ly resem­ble humans pro­voke dis­com­fort. Strik­ing the right bal­ance is para­mount, ensur­ing the AI’s respon­sive­ness aligns authen­ti­cal­ly with human expres­sions, with­out enter­ing the dis­com­fit­ing realm of the uncan­ny val­ley. The poten­tial of BCIs to fos­ter syn­chrony between the human brain and AI intro­duces promis­ing yet chal­leng­ing prospects for human-com­put­er collaboration.

Neu­ro­science not only illu­mi­nates the basis of bio­log­i­cal intel­li­gence but may also guide devel­op­ment of arti­fi­cial intel­li­gence². Con­sid­er­ing evo­lu­tion­ary con­straints like space and com­mu­ni­ca­tion effi­cien­cy, which have shaped the emer­gence of effi­cient sys­tems in nature, prompts explo­ration of embed­ding sim­i­lar con­straints in AI sys­tems, envi­sion­ing organ­i­cal­ly evolv­ing arti­fi­cial envi­ron­ments opti­mised for effi­cien­cy and envi­ron­men­tal sus­tain­abil­i­ty, the focus of research in so-called “neu­ro­mor­phic computing.” 

For exam­ple, oscil­la­to­ry neur­al activ­i­ty appears to boost com­mu­ni­ca­tion between dis­tant brain areas. The brain employs a theta-gam­ma rhythm to pack­age and trans­mit infor­ma­tion, sim­i­lar to a postal ser­vice, there­by enhanc­ing effi­cient data trans­mis­sion and retrieval³. This inter­play has been likened to an advanced data trans­mis­sion sys­tem, where low-fre­quen­cy alpha and beta brain waves sup­press neur­al activ­i­ty asso­ci­at­ed with pre­dictable stim­uli, allow­ing neu­rons in sen­so­ry regions to high­light unex­pect­ed stim­uli via high­er-fre­quen­cy gam­ma waves. Bas­tos et al.⁴ found that inhibito­ry pre­dic­tions car­ried by alpha/beta waves typ­i­cal­ly flow back­ward through deep­er cor­ti­cal lay­ers, while exci­ta­to­ry gam­ma waves con­vey­ing infor­ma­tion about nov­el stim­uli prop­a­gate for­ward through super­fi­cial layers.

Recent AI exper­i­ments, par­tic­u­lar­ly those involv­ing OpenAI’s GPT‑4, unveil intrigu­ing par­al­lels with evo­lu­tion­ary learning.

In the mam­malian brain, sharp wave rip­ples (SPW-Rs) exert wide­spread exci­ta­to­ry influ­ence through­out the cor­tex and mul­ti­ple sub­cor­ti­cal nuclei⁵. With­in these SPW-Rs, neu­ronal spik­ing is metic­u­lous­ly orches­trat­ed both tem­po­ral­ly and spa­tial­ly by interneu­rons, facil­i­tat­ing the con­densed reac­ti­va­tion of seg­ments from wak­ing neu­ronal sequences⁶. This orches­trat­ed activ­i­ty aids in the trans­mis­sion of com­pressed hip­pocam­pal rep­re­sen­ta­tions to dis­trib­uted cir­cuits, there­by rein­forc­ing the process of mem­o­ry con­sol­i­da­tion⁷.

Recent AI exper­i­ments, par­tic­u­lar­ly those involv­ing OpenAI’s GPT‑4, unveil intrigu­ing par­al­lels with evo­lu­tion­ary learn­ing. Unlike tra­di­tion­al task-ori­ent­ed train­ing, GPT‑4 learns from exten­sive datasets, refin­ing its respons­es based on the accu­mu­lat­ed ‘expe­ri­ences’ – fur­ther­more pat­tern recog­ni­tion by GPTs par­al­lels pat­tern recog­ni­tion by lay­ers of neu­rons in the brain. This approach mir­rors the adapt­abil­i­ty observed in nat­ur­al evo­lu­tion, where organ­isms refine their behav­iours over time to bet­ter res­onate with their environment.

Draw­ing inspi­ra­tion from the archi­tec­ture of the brain, neur­al net­works in AI are con­struct­ed with nodes organ­ised in lay­ers that respond to inputs and then gen­er­ate out­puts. In the realm of human neur­al syn­chrony research, inves­ti­gat­ing the role of oscil­la­tions has proven to be a piv­otal area of inter­est. High-fre­quen­cy oscil­la­to­ry neur­al activ­i­ty stands out as a cru­cial ele­ment, demon­strat­ing its abil­i­ty to facil­i­tate com­mu­ni­ca­tion between dis­tant brain areas. A par­tic­u­lar­ly intrigu­ing phe­nom­e­non in this con­text is the theta-gam­ma neur­al code, show­cas­ing how our brains employ a dis­tinc­tive method of ‘pack­ag­ing’ and ‘trans­mit­ting’ infor­ma­tion, rem­i­nis­cent of a postal ser­vice metic­u­lous­ly wrap­ping pack­ages for effi­cient deliv­ery. This neur­al ‘pack­ag­ing’ sys­tem orches­trates spe­cif­ic rhythms, akin to a coor­di­nat­ed dance, ensur­ing the stream­lined trans­mis­sion of infor­ma­tion, and it is encap­su­lat­ed in what is known as the theta-gam­ma rhythm.

This per­spec­tive aligns with the con­cept of “neu­ro­mor­phic com­put­ing,” where AI archi­tec­ture is based on neur­al cir­cuit­ry. The key advan­tage of neu­ro­mor­phic com­put­ing lies in its com­pu­ta­tion­al effi­cien­cy, address­ing the sig­nif­i­cant ener­gy con­sump­tion chal­lenges faced by tra­di­tion­al AI mod­els. The train­ing of large AI mod­els, such as those used in nat­ur­al lan­guage pro­cess­ing or image recog­ni­tion, can con­sume an exor­bi­tant amount of ener­gy. For instance, train­ing a sin­gle AI mod­el can emit as much car­bon diox­ide as five cars over their entire lifes­pan⁸. More­over, researchers at the Uni­ver­si­ty of Mass­a­chu­setts, Amherst, found that the car­bon foot­print of train­ing deep learn­ing mod­els has been dou­bling approx­i­mate­ly every 3.5 months, far out­pac­ing improve­ments in com­pu­ta­tion­al effi­cien­cy⁹.

Neu­ro­mor­phic com­put­ing offers a promis­ing alter­na­tive. By mim­ic­k­ing the archi­tec­ture of the human brain, neu­ro­mor­phic sys­tems aim to achieve high­er com­pu­ta­tion­al effi­cien­cy and low­er ener­gy con­sump­tion com­pared to con­ven­tion­al AI archi­tec­tures¹⁰. For exam­ple, IBM’s TrueNorth neu­ro­mor­phic chip has demon­strat­ed sig­nif­i­cant orders of mag­ni­tude in ener­gy effi­cien­cy com­pared to tra­di­tion­al CPUs and GPUs¹¹. Addi­tion­al­ly, neu­ro­mor­phic com­put­ing archi­tec­tures are inher­ent­ly suit­ed for low-pow­er, real-time pro­cess­ing tasks, mak­ing them ide­al for appli­ca­tions like edge com­put­ing and autonomous sys­tems, fur­ther con­tribut­ing to ener­gy sav­ings and envi­ron­men­tal sustainability.

In the realm of train­ing and skill devel­op­ment, syn­chro­nised AI has the poten­tial to per­son­alise learn­ing expe­ri­ences based on an employ­ee’s unique learn­ing curve, facil­i­tat­ing faster and more effec­tive skill acqui­si­tion. From a cus­tomer engage­ment stand­point, syn­chro­nised AI inter­faces might more pre­cise­ly under­stand and, in some cas­es, antic­i­pate user expec­ta­tions based on advanced behav­iour­al patterns.

For oper­a­tional effi­cien­cy, espe­cial­ly in sec­tors like man­u­fac­tur­ing or logis­tics, AI sys­tems work­ing in coor­di­na­tion with each oth­er can opti­mise process­es, reduce waste, and strength­en the sup­ply chain. This would lead to increased prof­itabil­i­ty, with an ever-met greater abil­i­ty for sus­tain­abil­i­ty con­sid­er­a­tions inte­grat­ed. In risk man­age­ment, syn­chro­nised AI sys­tems analysing vast datasets col­lab­o­ra­tive­ly might bet­ter pre­dict poten­tial risks or mar­ket down­turns, equip­ping busi­ness­es and oth­er organ­i­sa­tions to pre­pare or piv­ot before a cri­sis emerges to lim­it all relat­ed social and soci­etal impact. Like­wise, syn­chro­nised AI sys­tems could pro­vide insights for more effi­cient urban plan­ning and envi­ron­men­tal pro­tec­tion strate­gies. This could lead to bet­ter traf­fic man­age­ment, ener­gy con­ser­va­tion, and pol­lu­tion con­trol, enhanc­ing the qual­i­ty of life in urban areas.

In var­i­ous domains beyond busi­ness, deploy­ment of AI with a proso­cial ori­en­ta­tion holds immense poten­tial for the well-being of human­i­ty and the plan­et. Par­tic­u­lar­ly in health­care, syn­chro­ni­sa­tion between the human brain and AI sys­tems could ush­er in a rev­o­lu­tion­ary era for patient care and med­ical research. Recent stud­ies high­light the pos­i­tive impact of clin­i­cians syn­chro­nis­ing their move­ments with patients, there­by increas­ing trust, and reduc­ing pain. Extend­ing this con­cept to AI chat­bots or AI-enabled robot­ic care­givers that are syn­chro­nised with those under their ‘care’ holds the promise of enhanc­ing patient expe­ri­ence and improv­ing out­comes, as evi­denced by recent research indi­cat­ing that LLMs out­per­formed physi­cians in diag­nos­ing ill­ness­es, and patients pre­ferred their interaction.

In the edu­ca­tion­al domain, the inte­gra­tion of AI sys­tems with a focus on syn­chrony is equal­ly promis­ing. Research demon­strat­ed that syn­chro­nized brain waves in high school class­rooms were pre­dic­tive of high­er per­for­mance and hap­pi­ness among stu­dents¹². This study under­scores the sig­nif­i­cance of neur­al syn­chrony in the learn­ing envi­ron­ment. By lever­ag­ing AI tutor­ing sys­tems capa­ble of detect­ing and respond­ing to stu­dents’ cog­ni­tive states in real-time, edu­ca­tion tech­nol­o­gy can poten­tial­ly repli­cate the pos­i­tive out­comes observed in syn­chro­nised class­room set­tings.  Incor­po­ra­tion of AI sys­tems that res­onate with stu­dents’ brain states has the poten­tial to cre­ate a more con­ducive and effec­tive learn­ing atmos­phere, opti­miz­ing engage­ment and fos­ter­ing pos­i­tive learn­ing outcomes.

The excite­ment sur­round­ing the prospects of brain-to-machine and machine-to-machine syn­chrony brings with it a set of para­mount con­cerns that neces­si­tate scruti­ny and that are all but tech­ni­cal. Data pri­va­cy emerges as a crit­i­cal appre­hen­sion, giv­en the inti­mate nature of neur­al infor­ma­tion being processed by these sys­tems. The eth­i­cal dimen­sions of such syn­chro­ni­sa­tion, par­tic­u­lar­ly in the realm of AI deci­sion-mak­ing, present com­plex chal­lenges that require care­ful con­sid­er­a­tion¹³¹⁴.

Expand­ing on these con­cerns, two over­ar­ch­ing issues demand height­ened atten­tion. First­ly, the preser­va­tion of human auton­o­my stands as a foun­da­tion­al prin­ci­ple. As we delve into the era of brain-machine syn­chrony, it becomes imper­a­tive to ensure that indi­vid­u­als retain their abil­i­ty to make informed choic­es. Avoid­ing sce­nar­ios where indi­vid­u­als feel coerced or manip­u­lat­ed by tech­nol­o­gy is cru­cial in uphold­ing eth­i­cal standards.

Sec­ond­ly, the ques­tion of equi­ty in access to these tech­nolo­gies emerges as a press­ing mat­ter. Cur­rent­ly, such advanced tech­nolo­gies are often cost­ly and may not be acces­si­ble to all seg­ments of soci­ety. This rais­es con­cerns about exac­er­bat­ing exist­ing inequal­i­ties¹⁵. A sce­nario where only cer­tain priv­i­leged groups can har­ness the ben­e­fits of brain-machine syn­chrony might deep­en soci­etal divides. More­over, the lack of aware­ness about these tech­nolo­gies fur­ther com­pounds issues of equi­table access¹⁶.

The inte­gra­tion of AI with human cog­ni­tion marks the thresh­old of an unprece­dent­ed era, where machines not only repli­cate human intel­li­gence but also mir­ror intri­cate behav­iour­al pat­terns and emo­tions. The poten­tial syn­chro­ni­sa­tion of AI with human intent and emo­tion holds the promise of redefin­ing the nature of human-machine col­lab­o­ra­tion and, per­haps, even the essence of the human con­di­tion. The out­come of har­mon­is­ing humans and machines will sig­nif­i­cant­ly impact human­i­ty and the plan­et, con­tin­gent upon the guid­ing human aspi­ra­tions in this pur­suit, and open oppor­tu­ni­ties for an advanced human-cen­tered AI expe­ri­ence, in a “Fusion Mode”, as coined in the “Arti­fi­cial Integri­ty” con­cept. This rais­es a time­less ques­tion, rever­ber­at­ing through the course of human his­to­ry: what do we val­ue, and why?

A cru­cial point to empha­sise is that the impli­ca­tions of syn­chro­nis­ing humans and machines extend far beyond the realm of AI experts; it encom­pass­es every indi­vid­ual. This under­scores the neces­si­ty to raise aware­ness and engage the pub­lic at every stage of this trans­for­ma­tive jour­ney. As the devel­op­ment of AI pro­gress­es, it is essen­tial to ensure that the eth­i­cal, soci­etal, and exis­ten­tial dimen­sions are shaped by col­lec­tive val­ues and reflec­tions, avoid­ing uni­lat­er­al deci­sions by Big Tech that may not align with the broad­er inter­ests of human­i­ty. What hap­pens next shapes our indi­vid­ual and col­lec­tive future. Get­ting it right is our shared responsibility.