How open-source AI could modernise public services

Arti­fi­cial intel­li­gence (AI), and more specif­i­cal­ly large lan­guage mod­els (LLMs) rep­re­sent a major oppor­tu­ni­ty to trans­form pub­lic ser­vices. AI can be used in many areas to improve effi­cien­cy, the qual­i­ty of ser­vices pro­vid­ed to cit­i­zens and deci­sion-mak­ing. How­ev­er, the imple­men­ta­tion of AI in pub­lic ser­vices presents major chal­lenges. First, the cho­sen solu­tion must guar­an­tee fair treat­ment and trans­paren­cy of deci­sions and actions on a case or ensure respect for fun­da­men­tal rights through­out its use. In addi­tion, the rig­or­ous pro­tec­tion of per­son­al data, which is often sen­si­tive in the con­text of pub­lic ser­vices, is a sig­nif­i­cant secu­ri­ty issue. Final­ly, the trans­paren­cy of deci­sions is a major fac­tor in the trust placed in the solu­tions used and their accept­abil­i­ty to cit­i­zens. Thus, the use of a solu­tion offer­ing a high lev­el of trans­paren­cy is an asset in the imple­men­ta­tion and accep­tance of arti­fi­cial intel­li­gence solu­tions. But in view of the com­plex­i­ty of the sub­ject, the cri­te­ria for ensur­ing the expect­ed lev­el of trans­paren­cy are far from easy to define.

The major lan­guage mod­els are based on neur­al net­works trained on a very large amount of data. From a sequence of words, they sta­tis­ti­cal­ly deter­mine the word that best match­es the giv­en sequence. By apply­ing this prin­ci­ple recur­sive­ly, LLMs are able to pro­duce struc­tured texts, giv­ing the impres­sion that the machine is analysing and under­stand­ing the ques­tion asked.

The text pro­duced will there­fore depend on:

• the algo­rithms used, which will enable the mod­el to weigh the impor­tance of each word in a sen­tence in rela­tion to the oth­ers. This capac­i­ty is pro­vid­ed in par­tic­u­lar through “trans­former” type archi­tec­tures¹.
• the weight assigned to the dif­fer­ent neu­rons, which will enable the net­work to be acti­vat­ed in order to pro­duce the out­put data;
• the learn­ing cor­pus, which has a direct impact on the deter­mi­na­tion of the weights used by the model.

The 4 prin­ci­ples (use, study, mod­i­fy, share) asso­ci­at­ed with free soft­ware² must there­fore be applied to all these ele­ments³. This is still a sub­ject of debate and is there­fore a source of much con­fu­sion⁴. For exam­ple, some AIs claim­ing to be free have usage restric­tions that go against the defined prin­ci­ples⁵. After a long process, the Open Source Ini­tia­tive (OSI), which brings togeth­er researchers, lawyers, pol­i­cy mak­ers, activists and rep­re­sen­ta­tives of large tech­nol­o­gy com­pa­nies, has pro­posed a def­i­n­i­tion that cor­re­lates the four free­doms asso­ci­at­ed with free soft­ware and the ele­ments on which machine learn­ing mod­els are based.

Accord­ing to the Open Source Ini­tia­tive, a free machine learn­ing sys­tem must include the fol­low­ing ele­ments⁶:

• suf­fi­cient­ly detailed infor­ma­tion on the data used to train the sys­tem, enabling a com­pe­tent per­son to build a sub­stan­tial­ly equiv­a­lent sys­tem. This infor­ma­tion must be avail­able under terms approved by the OSI;
• the source code of the AI, includ­ing the infer­ence code to exe­cute the model;
• all the learned para­me­ters that are super­im­posed on the mod­el archi­tec­ture to pro­duce an out­put from a giv­en input.

The pub­li­ca­tion of the learn­ing cor­pus is there­fore not com­pul­so­ry, but a detailed descrip­tion of the lat­ter must be includ­ed. It is clear that many mod­els offer­ing excel­lent per­for­mance and describ­ing them­selves as open source do not com­ply with this last point. These are referred to as open-weight mod­els. A com­par­i­son of AI mod­els is also avail­able from the Pôle d’Ex­per­tise de la Régu­la­tion Numérique (PEReN).

The source code is human-read­able and pro­vides access to the algo­rithms used. The weights are the result of train­ing and rep­re­sent the knowl­edge of the mod­el. In the case of open-weight mod­els, this knowl­edge can be cus­tomised through a fine-tun­ing process⁷.

How­ev­er, this does not allow for total trans­paren­cy, such as the detec­tion of bias or “poi­son­ing” attacks, which con­sist of alter­ing the knowl­edge of a mod­el with­out these mod­i­fi­ca­tions being eas­i­ly detectable by stan­dard tests⁸⁹. Only a free mod­el that pro­vides access to its learn­ing cor­pus guar­an­tees a total lev­el of trans­paren­cy, in par­tic­u­lar by allow­ing com­plete con­trol of its train­ing. How­ev­er, this approach of recon­struct­ing from the sources still requires sig­nif­i­cant com­put­ing resources that few enti­ties are able to acquire.

On 30^th Octo­ber 2023, Pres­i­dent Biden issued an exec­u­tive order enti­tled Safe, Secure, and Trust­wor­thy Devel­op­ment and Use of Arti­fi­cial Intel­li­gence to assess the risks and ben­e­fits of foun­da­tion mod­els for which weights are avail­able. The report result­ing from this study¹⁰ recog­nis­es the ben­e­fits of open access to mod­el weights, such as inno­va­tion and research, but also high­lights the poten­tial risks, includ­ing the pos­si­bil­i­ty of mali­cious use, the removal of safe­ty mech­a­nisms and the impact on com­pe­ti­tion. The report con­cludes that the cur­rent data is insuf­fi­cient to defin­i­tive­ly deter­mine whether restric­tions on open-weight mod­els are jus­ti­fied and rec­om­mends active mon­i­tor­ing of these models.

Closed mod­els, even if they do not ben­e­fit from the same lev­el of trans­paren­cy and adapt­abil­i­ty as their free or open-weight coun­ter­parts, are not with­out advan­tages. They are less sub­ject to the risks of manip­u­la­tion men­tioned above because their weights can­not be mod­i­fied by a third par­ty, the risks to the intel­lec­tu­al prop­er­ty of the train­ing data are borne by the mod­el sup­pli­er, the pub­lish­er can quick­ly act on its mod­el in order to react in the event of abuse, thus help­ing to mit­i­gate the poten­tial risks asso­ci­at­ed with AI, such as the dis­sem­i­na­tion of inap­pro­pri­ate con­tent¹¹. How­ev­er, all this is at the expense of the auton­o­my that can be had over the AI model.

The use of open AIs as defined by the OSI has many advan­tages. First of all, the trans­paren­cy of their oper­a­tion is guar­an­teed, since it is pos­si­ble to direct­ly access and mod­i­fy their source code and inspect the train­ing data.

This pos­si­bil­i­ty is a fun­da­men­tal guar­an­tee since each mod­el used can be sub­ject­ed to in-depth ver­i­fi­ca­tion to ensure that the deci­sion-mak­ing process com­plies with cur­rent leg­is­la­tion and does not present any dis­crim­i­na­to­ry bias, for exam­ple. On the oth­er hand, when AI is used in the con­text of retrieval-aug­ment­ed gen­er­a­tion (RAG¹²), the lev­el of trans­paren­cy that must be required may be low­er because the data used to for­mu­late the respons­es are pro­vid­ed through an algo­rithm over which it is eas­i­er to have the expect­ed lev­el of con­trol. As the body of answers is pro­vid­ed by con­ven­tion­al search algo­rithms, it is rel­a­tive­ly easy to pro­vide the end user with the expect­ed answer, the raw data and their lev­el of con­fi­dence. How­ev­er, this requires a crit­i­cal eye on the part of the end user.

Even if the State’s mis­sions are rel­a­tive­ly spe­cif­ic by their very nature, there are many cas­es of use that are sim­i­lar to those that can be found in pri­vate com­pa­nies, name­ly pro­vid­ing an answer to a ques­tion by using a cor­pus of doc­u­ments with the help of clas­sic or vec­tor search algo­rithms based on the con­cept of sim­i­lar­i­ty¹³. It is there­fore not sur­pris­ing to see a con­ver­gence in the mod­els used in both worlds. For the State, the decid­ing cri­te­ri­on in the choice of mod­els will there­fore be relat­ed to the preser­va­tion of per­son­al infor­ma­tion or sen­si­tive infor­ma­tion trans­mit­ted to AI models.

The use of open-source solu­tions makes it pos­si­ble to dras­ti­cal­ly reduce expenses

Beyond the aspects men­tioned above, the use of open-source solu­tions also allows the State to dis­sem­i­nate its work so that it can be reused by the pub­lic or pri­vate sec­tor. For exam­ple, the DGFiP has pub­lished work on a mod­el for syn­the­sis­ing par­lia­men­tary amend­ments¹⁴¹⁵. They are thus able to active­ly share their knowl­edge with­in the lim­its of con­fi­den­tial­i­ty nec­es­sary for sov­er­eign missions.

Final­ly, the use of open-source solu­tions makes it pos­si­ble to dras­ti­cal­ly reduce expens­es, by lim­it­ing them to tech­ni­cal sup­port with­out licence costs.

The use of AI under a free licence also presents var­i­ous chal­lenges. First of all, the imple­men­ta­tion of free solu­tions requires a good under­stand­ing of how the under­ly­ing mod­els work. In addi­tion to this com­plex­i­ty, there is also the need for tech­ni­cal skills to adapt the mod­els to busi­ness needs, to have the data nec­es­sary for learn­ing, to con­fig­ure the mod­el (fine-tun­ing), if the busi­ness appli­ca­tion requires it, to deploy it in the admin­is­tra­tion’s IS and to guar­an­tee the high­est lev­el of security.

In addi­tion, their ongo­ing main­te­nance and cor­rec­tive main­te­nance require a sig­nif­i­cant invest­ment of time, both to update the mod­els or ensure a sat­is­fac­to­ry lev­el of non-regres­sion and to ensure that they func­tion prop­er­ly. Although the code is free, the use of these AIs often also requires IT infra­struc­tures based on spe­cialised com­put­ing units, which can rep­re­sent an indi­rect cost. Final­ly, the qual­i­ty of open-source mod­els can vary con­sid­er­ably, par­tic­u­lar­ly depend­ing on the busi­ness cas­es to be addressed, and there are no absolute guar­an­tees as to their per­for­mance. It is there­fore essen­tial to define the expec­ta­tions pre­cise­ly with the busi­ness teams and to ver­i­fy the expect­ed results before any ver­sion is put into service.

The inte­gra­tion of arti­fi­cial intel­li­gence with­in pub­lic ser­vices rep­re­sents a unique oppor­tu­ni­ty to improve the effi­cien­cy and qual­i­ty of ser­vices pro­vid­ed to cit­i­zens and deci­sion-mak­ing in a con­text of strain on avail­able human resources. Free lan­guage mod­els seem to be par­tic­u­lar­ly well-suit­ed to this challenge.

Despite the chal­lenges, the advan­tages of free AIs are numer­ous. They pro­mote inno­va­tion, reduce costs and strength­en the auton­o­my of administrations.

How­ev­er, it is essen­tial to study in depth the eth­i­cal issues relat­ed to the use of AI in the pub­lic sec­tor. It is nec­es­sary to put in place process­es and meth­ods to guard against algo­rith­mic bias and to guar­an­tee the rea­son­able use of tech­nolo­gies, ensur­ing that they are mon­i­tored by dig­i­tal and legal experts, or even by cit­i­zens themselves.

Dis­claimer: The con­tent of this arti­cle is the sole respon­si­bil­i­ty of its authors and has no scope oth­er than that of infor­ma­tion and aca­d­e­m­ic research.